1 //====- X86InstrSSE.td - Describe the X86 Instruction Set --*- tablegen -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file describes the X86 SSE instruction set, defining the instructions,
11 // and properties of the instructions which are needed for code generation,
12 // machine code emission, and analysis.
14 //===----------------------------------------------------------------------===//
17 //===----------------------------------------------------------------------===//
18 // SSE specific DAG Nodes.
19 //===----------------------------------------------------------------------===//
21 def SDTX86FPShiftOp : SDTypeProfile<1, 2, [ SDTCisSameAs<0, 1>,
22 SDTCisFP<0>, SDTCisInt<2> ]>;
23 def SDTX86VFCMP : SDTypeProfile<1, 3, [SDTCisInt<0>, SDTCisSameAs<1, 2>,
24 SDTCisFP<1>, SDTCisVT<3, i8>]>;
26 def X86fmin : SDNode<"X86ISD::FMIN", SDTFPBinOp>;
27 def X86fmax : SDNode<"X86ISD::FMAX", SDTFPBinOp>;
28 def X86fand : SDNode<"X86ISD::FAND", SDTFPBinOp,
29 [SDNPCommutative, SDNPAssociative]>;
30 def X86for : SDNode<"X86ISD::FOR", SDTFPBinOp,
31 [SDNPCommutative, SDNPAssociative]>;
32 def X86fxor : SDNode<"X86ISD::FXOR", SDTFPBinOp,
33 [SDNPCommutative, SDNPAssociative]>;
34 def X86frsqrt : SDNode<"X86ISD::FRSQRT", SDTFPUnaryOp>;
35 def X86frcp : SDNode<"X86ISD::FRCP", SDTFPUnaryOp>;
36 def X86fsrl : SDNode<"X86ISD::FSRL", SDTX86FPShiftOp>;
37 def X86comi : SDNode<"X86ISD::COMI", SDTX86CmpTest>;
38 def X86ucomi : SDNode<"X86ISD::UCOMI", SDTX86CmpTest>;
39 def X86pshufb : SDNode<"X86ISD::PSHUFB",
40 SDTypeProfile<1, 2, [SDTCisVT<0, v16i8>, SDTCisSameAs<0,1>,
42 def X86pextrb : SDNode<"X86ISD::PEXTRB",
43 SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisPtrTy<2>]>>;
44 def X86pextrw : SDNode<"X86ISD::PEXTRW",
45 SDTypeProfile<1, 2, [SDTCisVT<0, i32>, SDTCisPtrTy<2>]>>;
46 def X86pinsrb : SDNode<"X86ISD::PINSRB",
47 SDTypeProfile<1, 3, [SDTCisVT<0, v16i8>, SDTCisSameAs<0,1>,
48 SDTCisVT<2, i32>, SDTCisPtrTy<3>]>>;
49 def X86pinsrw : SDNode<"X86ISD::PINSRW",
50 SDTypeProfile<1, 3, [SDTCisVT<0, v8i16>, SDTCisSameAs<0,1>,
51 SDTCisVT<2, i32>, SDTCisPtrTy<3>]>>;
52 def X86insrtps : SDNode<"X86ISD::INSERTPS",
53 SDTypeProfile<1, 3, [SDTCisVT<0, v4f32>, SDTCisSameAs<0,1>,
54 SDTCisVT<2, v4f32>, SDTCisPtrTy<3>]>>;
55 def X86vzmovl : SDNode<"X86ISD::VZEXT_MOVL",
56 SDTypeProfile<1, 1, [SDTCisSameAs<0,1>]>>;
57 def X86vzload : SDNode<"X86ISD::VZEXT_LOAD", SDTLoad,
58 [SDNPHasChain, SDNPMayLoad]>;
59 def X86vshl : SDNode<"X86ISD::VSHL", SDTIntShiftOp>;
60 def X86vshr : SDNode<"X86ISD::VSRL", SDTIntShiftOp>;
61 def X86cmpps : SDNode<"X86ISD::CMPPS", SDTX86VFCMP>;
62 def X86cmppd : SDNode<"X86ISD::CMPPD", SDTX86VFCMP>;
63 def X86pcmpeqb : SDNode<"X86ISD::PCMPEQB", SDTIntBinOp, [SDNPCommutative]>;
64 def X86pcmpeqw : SDNode<"X86ISD::PCMPEQW", SDTIntBinOp, [SDNPCommutative]>;
65 def X86pcmpeqd : SDNode<"X86ISD::PCMPEQD", SDTIntBinOp, [SDNPCommutative]>;
66 def X86pcmpeqq : SDNode<"X86ISD::PCMPEQQ", SDTIntBinOp, [SDNPCommutative]>;
67 def X86pcmpgtb : SDNode<"X86ISD::PCMPGTB", SDTIntBinOp>;
68 def X86pcmpgtw : SDNode<"X86ISD::PCMPGTW", SDTIntBinOp>;
69 def X86pcmpgtd : SDNode<"X86ISD::PCMPGTD", SDTIntBinOp>;
70 def X86pcmpgtq : SDNode<"X86ISD::PCMPGTQ", SDTIntBinOp>;
72 def SDTX86CmpPTest : SDTypeProfile<1, 2, [SDTCisVT<0, i32>,
75 def X86ptest : SDNode<"X86ISD::PTEST", SDTX86CmpPTest>;
77 //===----------------------------------------------------------------------===//
78 // SSE Complex Patterns
79 //===----------------------------------------------------------------------===//
81 // These are 'extloads' from a scalar to the low element of a vector, zeroing
82 // the top elements. These are used for the SSE 'ss' and 'sd' instruction
84 def sse_load_f32 : ComplexPattern<v4f32, 5, "SelectScalarSSELoad", [],
85 [SDNPHasChain, SDNPMayLoad]>;
86 def sse_load_f64 : ComplexPattern<v2f64, 5, "SelectScalarSSELoad", [],
87 [SDNPHasChain, SDNPMayLoad]>;
89 def ssmem : Operand<v4f32> {
90 let PrintMethod = "printf32mem";
91 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc_nosp, i32imm, i8imm);
92 let ParserMatchClass = X86MemAsmOperand;
94 def sdmem : Operand<v2f64> {
95 let PrintMethod = "printf64mem";
96 let MIOperandInfo = (ops ptr_rc, i8imm, ptr_rc_nosp, i32imm, i8imm);
97 let ParserMatchClass = X86MemAsmOperand;
100 //===----------------------------------------------------------------------===//
101 // SSE pattern fragments
102 //===----------------------------------------------------------------------===//
104 def loadv4f32 : PatFrag<(ops node:$ptr), (v4f32 (load node:$ptr))>;
105 def loadv2f64 : PatFrag<(ops node:$ptr), (v2f64 (load node:$ptr))>;
106 def loadv4i32 : PatFrag<(ops node:$ptr), (v4i32 (load node:$ptr))>;
107 def loadv2i64 : PatFrag<(ops node:$ptr), (v2i64 (load node:$ptr))>;
109 // Like 'store', but always requires vector alignment.
110 def alignedstore : PatFrag<(ops node:$val, node:$ptr),
111 (store node:$val, node:$ptr), [{
112 return cast<StoreSDNode>(N)->getAlignment() >= 16;
115 // Like 'load', but always requires vector alignment.
116 def alignedload : PatFrag<(ops node:$ptr), (load node:$ptr), [{
117 return cast<LoadSDNode>(N)->getAlignment() >= 16;
120 def alignedloadfsf32 : PatFrag<(ops node:$ptr),
121 (f32 (alignedload node:$ptr))>;
122 def alignedloadfsf64 : PatFrag<(ops node:$ptr),
123 (f64 (alignedload node:$ptr))>;
124 def alignedloadv4f32 : PatFrag<(ops node:$ptr),
125 (v4f32 (alignedload node:$ptr))>;
126 def alignedloadv2f64 : PatFrag<(ops node:$ptr),
127 (v2f64 (alignedload node:$ptr))>;
128 def alignedloadv4i32 : PatFrag<(ops node:$ptr),
129 (v4i32 (alignedload node:$ptr))>;
130 def alignedloadv2i64 : PatFrag<(ops node:$ptr),
131 (v2i64 (alignedload node:$ptr))>;
133 // Like 'load', but uses special alignment checks suitable for use in
134 // memory operands in most SSE instructions, which are required to
135 // be naturally aligned on some targets but not on others. If the subtarget
136 // allows unaligned accesses, match any load, though this may require
137 // setting a feature bit in the processor (on startup, for example).
138 // Opteron 10h and later implement such a feature.
139 def memop : PatFrag<(ops node:$ptr), (load node:$ptr), [{
140 return Subtarget->hasVectorUAMem()
141 || cast<LoadSDNode>(N)->getAlignment() >= 16;
144 def memopfsf32 : PatFrag<(ops node:$ptr), (f32 (memop node:$ptr))>;
145 def memopfsf64 : PatFrag<(ops node:$ptr), (f64 (memop node:$ptr))>;
146 def memopv4f32 : PatFrag<(ops node:$ptr), (v4f32 (memop node:$ptr))>;
147 def memopv2f64 : PatFrag<(ops node:$ptr), (v2f64 (memop node:$ptr))>;
148 def memopv4i32 : PatFrag<(ops node:$ptr), (v4i32 (memop node:$ptr))>;
149 def memopv2i64 : PatFrag<(ops node:$ptr), (v2i64 (memop node:$ptr))>;
150 def memopv16i8 : PatFrag<(ops node:$ptr), (v16i8 (memop node:$ptr))>;
152 // SSSE3 uses MMX registers for some instructions. They aren't aligned on a
154 // FIXME: 8 byte alignment for mmx reads is not required
155 def memop64 : PatFrag<(ops node:$ptr), (load node:$ptr), [{
156 return cast<LoadSDNode>(N)->getAlignment() >= 8;
159 def memopv8i8 : PatFrag<(ops node:$ptr), (v8i8 (memop64 node:$ptr))>;
160 def memopv4i16 : PatFrag<(ops node:$ptr), (v4i16 (memop64 node:$ptr))>;
161 def memopv8i16 : PatFrag<(ops node:$ptr), (v8i16 (memop64 node:$ptr))>;
162 def memopv2i32 : PatFrag<(ops node:$ptr), (v2i32 (memop64 node:$ptr))>;
165 // Like 'store', but requires the non-temporal bit to be set
166 def nontemporalstore : PatFrag<(ops node:$val, node:$ptr),
167 (st node:$val, node:$ptr), [{
168 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
169 return ST->isNonTemporal();
173 def alignednontemporalstore : PatFrag<(ops node:$val, node:$ptr),
174 (st node:$val, node:$ptr), [{
175 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
176 return ST->isNonTemporal() && !ST->isTruncatingStore() &&
177 ST->getAddressingMode() == ISD::UNINDEXED &&
178 ST->getAlignment() >= 16;
182 def unalignednontemporalstore : PatFrag<(ops node:$val, node:$ptr),
183 (st node:$val, node:$ptr), [{
184 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N))
185 return ST->isNonTemporal() &&
186 ST->getAlignment() < 16;
190 def bc_v4f32 : PatFrag<(ops node:$in), (v4f32 (bitconvert node:$in))>;
191 def bc_v2f64 : PatFrag<(ops node:$in), (v2f64 (bitconvert node:$in))>;
192 def bc_v16i8 : PatFrag<(ops node:$in), (v16i8 (bitconvert node:$in))>;
193 def bc_v8i16 : PatFrag<(ops node:$in), (v8i16 (bitconvert node:$in))>;
194 def bc_v4i32 : PatFrag<(ops node:$in), (v4i32 (bitconvert node:$in))>;
195 def bc_v2i64 : PatFrag<(ops node:$in), (v2i64 (bitconvert node:$in))>;
197 def vzmovl_v2i64 : PatFrag<(ops node:$src),
198 (bitconvert (v2i64 (X86vzmovl
199 (v2i64 (scalar_to_vector (loadi64 node:$src))))))>;
200 def vzmovl_v4i32 : PatFrag<(ops node:$src),
201 (bitconvert (v4i32 (X86vzmovl
202 (v4i32 (scalar_to_vector (loadi32 node:$src))))))>;
204 def vzload_v2i64 : PatFrag<(ops node:$src),
205 (bitconvert (v2i64 (X86vzload node:$src)))>;
208 def fp32imm0 : PatLeaf<(f32 fpimm), [{
209 return N->isExactlyValue(+0.0);
212 // BYTE_imm - Transform bit immediates into byte immediates.
213 def BYTE_imm : SDNodeXForm<imm, [{
214 // Transformation function: imm >> 3
215 return getI32Imm(N->getZExtValue() >> 3);
218 // SHUFFLE_get_shuf_imm xform function: convert vector_shuffle mask to PSHUF*,
220 def SHUFFLE_get_shuf_imm : SDNodeXForm<vector_shuffle, [{
221 return getI8Imm(X86::getShuffleSHUFImmediate(N));
224 // SHUFFLE_get_pshufhw_imm xform function: convert vector_shuffle mask to
226 def SHUFFLE_get_pshufhw_imm : SDNodeXForm<vector_shuffle, [{
227 return getI8Imm(X86::getShufflePSHUFHWImmediate(N));
230 // SHUFFLE_get_pshuflw_imm xform function: convert vector_shuffle mask to
232 def SHUFFLE_get_pshuflw_imm : SDNodeXForm<vector_shuffle, [{
233 return getI8Imm(X86::getShufflePSHUFLWImmediate(N));
236 // SHUFFLE_get_palign_imm xform function: convert vector_shuffle mask to
238 def SHUFFLE_get_palign_imm : SDNodeXForm<vector_shuffle, [{
239 return getI8Imm(X86::getShufflePALIGNRImmediate(N));
242 def splat_lo : PatFrag<(ops node:$lhs, node:$rhs),
243 (vector_shuffle node:$lhs, node:$rhs), [{
244 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
245 return SVOp->isSplat() && SVOp->getSplatIndex() == 0;
248 def movddup : PatFrag<(ops node:$lhs, node:$rhs),
249 (vector_shuffle node:$lhs, node:$rhs), [{
250 return X86::isMOVDDUPMask(cast<ShuffleVectorSDNode>(N));
253 def movhlps : PatFrag<(ops node:$lhs, node:$rhs),
254 (vector_shuffle node:$lhs, node:$rhs), [{
255 return X86::isMOVHLPSMask(cast<ShuffleVectorSDNode>(N));
258 def movhlps_undef : PatFrag<(ops node:$lhs, node:$rhs),
259 (vector_shuffle node:$lhs, node:$rhs), [{
260 return X86::isMOVHLPS_v_undef_Mask(cast<ShuffleVectorSDNode>(N));
263 def movlhps : PatFrag<(ops node:$lhs, node:$rhs),
264 (vector_shuffle node:$lhs, node:$rhs), [{
265 return X86::isMOVLHPSMask(cast<ShuffleVectorSDNode>(N));
268 def movlp : PatFrag<(ops node:$lhs, node:$rhs),
269 (vector_shuffle node:$lhs, node:$rhs), [{
270 return X86::isMOVLPMask(cast<ShuffleVectorSDNode>(N));
273 def movl : PatFrag<(ops node:$lhs, node:$rhs),
274 (vector_shuffle node:$lhs, node:$rhs), [{
275 return X86::isMOVLMask(cast<ShuffleVectorSDNode>(N));
278 def movshdup : PatFrag<(ops node:$lhs, node:$rhs),
279 (vector_shuffle node:$lhs, node:$rhs), [{
280 return X86::isMOVSHDUPMask(cast<ShuffleVectorSDNode>(N));
283 def movsldup : PatFrag<(ops node:$lhs, node:$rhs),
284 (vector_shuffle node:$lhs, node:$rhs), [{
285 return X86::isMOVSLDUPMask(cast<ShuffleVectorSDNode>(N));
288 def unpckl : PatFrag<(ops node:$lhs, node:$rhs),
289 (vector_shuffle node:$lhs, node:$rhs), [{
290 return X86::isUNPCKLMask(cast<ShuffleVectorSDNode>(N));
293 def unpckh : PatFrag<(ops node:$lhs, node:$rhs),
294 (vector_shuffle node:$lhs, node:$rhs), [{
295 return X86::isUNPCKHMask(cast<ShuffleVectorSDNode>(N));
298 def unpckl_undef : PatFrag<(ops node:$lhs, node:$rhs),
299 (vector_shuffle node:$lhs, node:$rhs), [{
300 return X86::isUNPCKL_v_undef_Mask(cast<ShuffleVectorSDNode>(N));
303 def unpckh_undef : PatFrag<(ops node:$lhs, node:$rhs),
304 (vector_shuffle node:$lhs, node:$rhs), [{
305 return X86::isUNPCKH_v_undef_Mask(cast<ShuffleVectorSDNode>(N));
308 def pshufd : PatFrag<(ops node:$lhs, node:$rhs),
309 (vector_shuffle node:$lhs, node:$rhs), [{
310 return X86::isPSHUFDMask(cast<ShuffleVectorSDNode>(N));
311 }], SHUFFLE_get_shuf_imm>;
313 def shufp : PatFrag<(ops node:$lhs, node:$rhs),
314 (vector_shuffle node:$lhs, node:$rhs), [{
315 return X86::isSHUFPMask(cast<ShuffleVectorSDNode>(N));
316 }], SHUFFLE_get_shuf_imm>;
318 def pshufhw : PatFrag<(ops node:$lhs, node:$rhs),
319 (vector_shuffle node:$lhs, node:$rhs), [{
320 return X86::isPSHUFHWMask(cast<ShuffleVectorSDNode>(N));
321 }], SHUFFLE_get_pshufhw_imm>;
323 def pshuflw : PatFrag<(ops node:$lhs, node:$rhs),
324 (vector_shuffle node:$lhs, node:$rhs), [{
325 return X86::isPSHUFLWMask(cast<ShuffleVectorSDNode>(N));
326 }], SHUFFLE_get_pshuflw_imm>;
328 def palign : PatFrag<(ops node:$lhs, node:$rhs),
329 (vector_shuffle node:$lhs, node:$rhs), [{
330 return X86::isPALIGNRMask(cast<ShuffleVectorSDNode>(N));
331 }], SHUFFLE_get_palign_imm>;
333 //===----------------------------------------------------------------------===//
334 // SSE scalar FP Instructions
335 //===----------------------------------------------------------------------===//
337 // CMOV* - Used to implement the SSE SELECT DAG operation. Expanded after
338 // instruction selection into a branch sequence.
339 let Uses = [EFLAGS], usesCustomInserter = 1 in {
340 def CMOV_FR32 : I<0, Pseudo,
341 (outs FR32:$dst), (ins FR32:$t, FR32:$f, i8imm:$cond),
342 "#CMOV_FR32 PSEUDO!",
343 [(set FR32:$dst, (X86cmov FR32:$t, FR32:$f, imm:$cond,
345 def CMOV_FR64 : I<0, Pseudo,
346 (outs FR64:$dst), (ins FR64:$t, FR64:$f, i8imm:$cond),
347 "#CMOV_FR64 PSEUDO!",
348 [(set FR64:$dst, (X86cmov FR64:$t, FR64:$f, imm:$cond,
350 def CMOV_V4F32 : I<0, Pseudo,
351 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
352 "#CMOV_V4F32 PSEUDO!",
354 (v4f32 (X86cmov VR128:$t, VR128:$f, imm:$cond,
356 def CMOV_V2F64 : I<0, Pseudo,
357 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
358 "#CMOV_V2F64 PSEUDO!",
360 (v2f64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
362 def CMOV_V2I64 : I<0, Pseudo,
363 (outs VR128:$dst), (ins VR128:$t, VR128:$f, i8imm:$cond),
364 "#CMOV_V2I64 PSEUDO!",
366 (v2i64 (X86cmov VR128:$t, VR128:$f, imm:$cond,
370 //===----------------------------------------------------------------------===//
372 //===----------------------------------------------------------------------===//
374 // Move Instructions. Register-to-register movss is not used for FR32
375 // register copies because it's a partial register update; FsMOVAPSrr is
376 // used instead. Register-to-register movss is not modeled as an INSERT_SUBREG
377 // because INSERT_SUBREG requires that the insert be implementable in terms of
378 // a copy, and just mentioned, we don't use movss for copies.
379 let Constraints = "$src1 = $dst" in
380 def MOVSSrr : SSI<0x10, MRMSrcReg,
381 (outs VR128:$dst), (ins VR128:$src1, FR32:$src2),
382 "movss\t{$src2, $dst|$dst, $src2}",
383 [(set (v4f32 VR128:$dst),
384 (movl VR128:$src1, (scalar_to_vector FR32:$src2)))]>;
386 // Extract the low 32-bit value from one vector and insert it into another.
387 let AddedComplexity = 15 in
388 def : Pat<(v4f32 (movl VR128:$src1, VR128:$src2)),
389 (MOVSSrr (v4f32 VR128:$src1),
390 (EXTRACT_SUBREG (v4f32 VR128:$src2), sub_ss))>;
392 // Implicitly promote a 32-bit scalar to a vector.
393 def : Pat<(v4f32 (scalar_to_vector FR32:$src)),
394 (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), FR32:$src, sub_ss)>;
396 // Loading from memory automatically zeroing upper bits.
397 let canFoldAsLoad = 1, isReMaterializable = 1 in
398 def MOVSSrm : SSI<0x10, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
399 "movss\t{$src, $dst|$dst, $src}",
400 [(set FR32:$dst, (loadf32 addr:$src))]>;
402 // MOVSSrm zeros the high parts of the register; represent this
403 // with SUBREG_TO_REG.
404 let AddedComplexity = 20 in {
405 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector (loadf32 addr:$src))))),
406 (SUBREG_TO_REG (i32 0), (MOVSSrm addr:$src), sub_ss)>;
407 def : Pat<(v4f32 (scalar_to_vector (loadf32 addr:$src))),
408 (SUBREG_TO_REG (i32 0), (MOVSSrm addr:$src), sub_ss)>;
409 def : Pat<(v4f32 (X86vzmovl (loadv4f32 addr:$src))),
410 (SUBREG_TO_REG (i32 0), (MOVSSrm addr:$src), sub_ss)>;
413 // Store scalar value to memory.
414 def MOVSSmr : SSI<0x11, MRMDestMem, (outs), (ins f32mem:$dst, FR32:$src),
415 "movss\t{$src, $dst|$dst, $src}",
416 [(store FR32:$src, addr:$dst)]>;
418 // Extract and store.
419 def : Pat<(store (f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
422 (EXTRACT_SUBREG (v4f32 VR128:$src), sub_ss))>;
424 // Conversion instructions
425 def CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR32:$src),
426 "cvttss2si\t{$src, $dst|$dst, $src}",
427 [(set GR32:$dst, (fp_to_sint FR32:$src))]>;
428 def CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
429 "cvttss2si\t{$src, $dst|$dst, $src}",
430 [(set GR32:$dst, (fp_to_sint (loadf32 addr:$src)))]>;
431 def CVTSI2SSrr : SSI<0x2A, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
432 "cvtsi2ss\t{$src, $dst|$dst, $src}",
433 [(set FR32:$dst, (sint_to_fp GR32:$src))]>;
434 def CVTSI2SSrm : SSI<0x2A, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
435 "cvtsi2ss\t{$src, $dst|$dst, $src}",
436 [(set FR32:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
438 // Match intrinsics which expect XMM operand(s).
439 def CVTSS2SIrr: SSI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins FR32:$src),
440 "cvtss2si{l}\t{$src, $dst|$dst, $src}", []>;
441 def CVTSS2SIrm: SSI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
442 "cvtss2si{l}\t{$src, $dst|$dst, $src}", []>;
444 def Int_CVTSS2SIrr : SSI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
445 "cvtss2si\t{$src, $dst|$dst, $src}",
446 [(set GR32:$dst, (int_x86_sse_cvtss2si VR128:$src))]>;
447 def Int_CVTSS2SIrm : SSI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
448 "cvtss2si\t{$src, $dst|$dst, $src}",
449 [(set GR32:$dst, (int_x86_sse_cvtss2si
450 (load addr:$src)))]>;
452 // Match intrinsics which expect MM and XMM operand(s).
453 def Int_CVTPS2PIrr : PSI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
454 "cvtps2pi\t{$src, $dst|$dst, $src}",
455 [(set VR64:$dst, (int_x86_sse_cvtps2pi VR128:$src))]>;
456 def Int_CVTPS2PIrm : PSI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
457 "cvtps2pi\t{$src, $dst|$dst, $src}",
458 [(set VR64:$dst, (int_x86_sse_cvtps2pi
459 (load addr:$src)))]>;
460 def Int_CVTTPS2PIrr: PSI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
461 "cvttps2pi\t{$src, $dst|$dst, $src}",
462 [(set VR64:$dst, (int_x86_sse_cvttps2pi VR128:$src))]>;
463 def Int_CVTTPS2PIrm: PSI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f64mem:$src),
464 "cvttps2pi\t{$src, $dst|$dst, $src}",
465 [(set VR64:$dst, (int_x86_sse_cvttps2pi
466 (load addr:$src)))]>;
467 let Constraints = "$src1 = $dst" in {
468 def Int_CVTPI2PSrr : PSI<0x2A, MRMSrcReg,
469 (outs VR128:$dst), (ins VR128:$src1, VR64:$src2),
470 "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
471 [(set VR128:$dst, (int_x86_sse_cvtpi2ps VR128:$src1,
473 def Int_CVTPI2PSrm : PSI<0x2A, MRMSrcMem,
474 (outs VR128:$dst), (ins VR128:$src1, i64mem:$src2),
475 "cvtpi2ps\t{$src2, $dst|$dst, $src2}",
476 [(set VR128:$dst, (int_x86_sse_cvtpi2ps VR128:$src1,
477 (load addr:$src2)))]>;
480 // Aliases for intrinsics
481 def Int_CVTTSS2SIrr : SSI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
482 "cvttss2si\t{$src, $dst|$dst, $src}",
484 (int_x86_sse_cvttss2si VR128:$src))]>;
485 def Int_CVTTSS2SIrm : SSI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f32mem:$src),
486 "cvttss2si\t{$src, $dst|$dst, $src}",
488 (int_x86_sse_cvttss2si(load addr:$src)))]>;
490 let Constraints = "$src1 = $dst" in {
491 def Int_CVTSI2SSrr : SSI<0x2A, MRMSrcReg,
492 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
493 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
494 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
496 def Int_CVTSI2SSrm : SSI<0x2A, MRMSrcMem,
497 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
498 "cvtsi2ss\t{$src2, $dst|$dst, $src2}",
499 [(set VR128:$dst, (int_x86_sse_cvtsi2ss VR128:$src1,
500 (loadi32 addr:$src2)))]>;
503 // Comparison instructions
504 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
505 def CMPSSrr : SSIi8<0xC2, MRMSrcReg,
506 (outs FR32:$dst), (ins FR32:$src1, FR32:$src, SSECC:$cc),
507 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
509 def CMPSSrm : SSIi8<0xC2, MRMSrcMem,
510 (outs FR32:$dst), (ins FR32:$src1, f32mem:$src, SSECC:$cc),
511 "cmp${cc}ss\t{$src, $dst|$dst, $src}", []>;
513 // Accept explicit immediate argument form instead of comparison code.
514 let isAsmParserOnly = 1 in {
515 def CMPSSrr_alt : SSIi8<0xC2, MRMSrcReg,
516 (outs FR32:$dst), (ins FR32:$src1, FR32:$src, i8imm:$src2),
517 "cmpss\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
519 def CMPSSrm_alt : SSIi8<0xC2, MRMSrcMem,
520 (outs FR32:$dst), (ins FR32:$src1, f32mem:$src, i8imm:$src2),
521 "cmpss\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
525 let Defs = [EFLAGS] in {
526 def UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins FR32:$src1, FR32:$src2),
527 "ucomiss\t{$src2, $src1|$src1, $src2}",
528 [(set EFLAGS, (X86cmp FR32:$src1, FR32:$src2))]>;
529 def UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs), (ins FR32:$src1, f32mem:$src2),
530 "ucomiss\t{$src2, $src1|$src1, $src2}",
531 [(set EFLAGS, (X86cmp FR32:$src1, (loadf32 addr:$src2)))]>;
533 def COMISSrr: PSI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
534 "comiss\t{$src2, $src1|$src1, $src2}", []>;
535 def COMISSrm: PSI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
536 "comiss\t{$src2, $src1|$src1, $src2}", []>;
540 // Aliases to match intrinsics which expect XMM operand(s).
541 let Constraints = "$src1 = $dst" in {
542 def Int_CMPSSrr : SSIi8<0xC2, MRMSrcReg,
544 (ins VR128:$src1, VR128:$src, SSECC:$cc),
545 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
546 [(set VR128:$dst, (int_x86_sse_cmp_ss
548 VR128:$src, imm:$cc))]>;
549 def Int_CMPSSrm : SSIi8<0xC2, MRMSrcMem,
551 (ins VR128:$src1, f32mem:$src, SSECC:$cc),
552 "cmp${cc}ss\t{$src, $dst|$dst, $src}",
553 [(set VR128:$dst, (int_x86_sse_cmp_ss VR128:$src1,
554 (load addr:$src), imm:$cc))]>;
557 let Defs = [EFLAGS] in {
558 def Int_UCOMISSrr: PSI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
559 "ucomiss\t{$src2, $src1|$src1, $src2}",
560 [(set EFLAGS, (X86ucomi (v4f32 VR128:$src1),
562 def Int_UCOMISSrm: PSI<0x2E, MRMSrcMem, (outs),(ins VR128:$src1, f128mem:$src2),
563 "ucomiss\t{$src2, $src1|$src1, $src2}",
564 [(set EFLAGS, (X86ucomi (v4f32 VR128:$src1),
565 (load addr:$src2)))]>;
567 def Int_COMISSrr: PSI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
568 "comiss\t{$src2, $src1|$src1, $src2}",
569 [(set EFLAGS, (X86comi (v4f32 VR128:$src1),
571 def Int_COMISSrm: PSI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
572 "comiss\t{$src2, $src1|$src1, $src2}",
573 [(set EFLAGS, (X86comi (v4f32 VR128:$src1),
574 (load addr:$src2)))]>;
577 // Aliases of packed SSE1 instructions for scalar use. These all have names
578 // that start with 'Fs'.
580 // Alias instructions that map fld0 to pxor for sse.
581 let isReMaterializable = 1, isAsCheapAsAMove = 1, isCodeGenOnly = 1,
583 // FIXME: Set encoding to pseudo!
584 def FsFLD0SS : I<0xEF, MRMInitReg, (outs FR32:$dst), (ins), "",
585 [(set FR32:$dst, fp32imm0)]>,
586 Requires<[HasSSE1]>, TB, OpSize;
588 // Alias instruction to do FR32 reg-to-reg copy using movaps. Upper bits are
590 let neverHasSideEffects = 1 in
591 def FsMOVAPSrr : PSI<0x28, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
592 "movaps\t{$src, $dst|$dst, $src}", []>;
594 // Alias instruction to load FR32 from f128mem using movaps. Upper bits are
596 let canFoldAsLoad = 1, isReMaterializable = 1 in
597 def FsMOVAPSrm : PSI<0x28, MRMSrcMem, (outs FR32:$dst), (ins f128mem:$src),
598 "movaps\t{$src, $dst|$dst, $src}",
599 [(set FR32:$dst, (alignedloadfsf32 addr:$src))]>;
601 /// sse12_fp_alias_pack_logical - SSE 1 & 2 aliased packed FP logical ops
603 multiclass sse12_fp_alias_pack_logical<bits<8> opc, string OpcodeStr,
604 SDNode OpNode, int NoPat = 0,
605 bit MayLoad = 0, bit Commutable = 1> {
606 def PSrr : PSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
607 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
609 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))])> {
610 let isCommutable = Commutable;
613 def PDrr : PDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
614 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
616 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))])> {
617 let isCommutable = Commutable;
620 def PSrm : PSI<opc, MRMSrcMem, (outs FR32:$dst),
621 (ins FR32:$src1, f128mem:$src2),
622 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
624 [(set FR32:$dst, (OpNode FR32:$src1,
625 (memopfsf32 addr:$src2)))])> {
626 let mayLoad = MayLoad;
629 def PDrm : PDI<opc, MRMSrcMem, (outs FR64:$dst),
630 (ins FR64:$src1, f128mem:$src2),
631 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
633 [(set FR64:$dst, (OpNode FR64:$src1,
634 (memopfsf64 addr:$src2)))])> {
635 let mayLoad = MayLoad;
639 // Alias bitwise logical operations using SSE logical ops on packed FP values.
640 let Constraints = "$src1 = $dst" in {
641 defm FsAND : sse12_fp_alias_pack_logical<0x54, "and", X86fand>;
642 defm FsOR : sse12_fp_alias_pack_logical<0x56, "or", X86for>;
643 defm FsXOR : sse12_fp_alias_pack_logical<0x57, "xor", X86fxor>;
645 let neverHasSideEffects = 1 in
646 defm FsANDN : sse12_fp_alias_pack_logical<0x55, "andn", undef, 1, 1, 0>;
649 /// basic_sse12_fp_binop_rm - SSE 1 & 2 binops come in both scalar and
652 /// In addition, we also have a special variant of the scalar form here to
653 /// represent the associated intrinsic operation. This form is unlike the
654 /// plain scalar form, in that it takes an entire vector (instead of a scalar)
655 /// and leaves the top elements unmodified (therefore these cannot be commuted).
657 /// These three forms can each be reg+reg or reg+mem, so there are a total of
658 /// six "instructions".
660 let Constraints = "$src1 = $dst" in {
661 multiclass basic_sse12_fp_binop_rm<bits<8> opc, string OpcodeStr,
662 SDNode OpNode, bit Commutable = 0> {
663 // Scalar operation, reg+reg.
664 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
665 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
666 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
667 let isCommutable = Commutable;
670 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
671 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
672 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
673 let isCommutable = Commutable;
676 // Scalar operation, reg+mem.
677 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst),
678 (ins FR32:$src1, f32mem:$src2),
679 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
680 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
682 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst),
683 (ins FR64:$src1, f64mem:$src2),
684 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
685 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
687 // Vector operation, reg+reg.
688 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst),
689 (ins VR128:$src1, VR128:$src2),
690 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
691 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
692 let isCommutable = Commutable;
695 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
696 (ins VR128:$src1, VR128:$src2),
697 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
698 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
699 let isCommutable = Commutable;
702 // Vector operation, reg+mem.
703 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst),
704 (ins VR128:$src1, f128mem:$src2),
705 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
706 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
708 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
709 (ins VR128:$src1, f128mem:$src2),
710 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
711 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
713 // Intrinsic operation, reg+reg.
714 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst),
715 (ins VR128:$src1, VR128:$src2),
716 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
717 [(set VR128:$dst, (!nameconcat<Intrinsic>("int_x86_sse_",
718 !strconcat(OpcodeStr, "_ss")) VR128:$src1,
720 // int_x86_sse_xxx_ss
722 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst),
723 (ins VR128:$src1, VR128:$src2),
724 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
725 [(set VR128:$dst, (!nameconcat<Intrinsic>("int_x86_sse2_",
726 !strconcat(OpcodeStr, "_sd")) VR128:$src1,
728 // int_x86_sse2_xxx_sd
730 // Intrinsic operation, reg+mem.
731 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst),
732 (ins VR128:$src1, ssmem:$src2),
733 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
734 [(set VR128:$dst, (!nameconcat<Intrinsic>("int_x86_sse_",
735 !strconcat(OpcodeStr, "_ss")) VR128:$src1,
736 sse_load_f32:$src2))]>;
737 // int_x86_sse_xxx_ss
739 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst),
740 (ins VR128:$src1, sdmem:$src2),
741 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
742 [(set VR128:$dst, (!nameconcat<Intrinsic>("int_x86_sse2_",
743 !strconcat(OpcodeStr, "_sd")) VR128:$src1,
744 sse_load_f64:$src2))]>;
745 // int_x86_sse2_xxx_sd
749 // Arithmetic instructions
750 defm ADD : basic_sse12_fp_binop_rm<0x58, "add", fadd, 1>;
751 defm MUL : basic_sse12_fp_binop_rm<0x59, "mul", fmul, 1>;
752 defm SUB : basic_sse12_fp_binop_rm<0x5C, "sub", fsub>;
753 defm DIV : basic_sse12_fp_binop_rm<0x5E, "div", fdiv>;
755 /// sse12_fp_binop_rm - Other SSE 1 & 2 binops
757 /// This multiclass is like basic_sse12_fp_binop_rm, with the addition of
758 /// instructions for a full-vector intrinsic form. Operations that map
759 /// onto C operators don't use this form since they just use the plain
760 /// vector form instead of having a separate vector intrinsic form.
762 /// This provides a total of eight "instructions".
764 let Constraints = "$src1 = $dst" in {
765 multiclass sse12_fp_binop_rm<bits<8> opc, string OpcodeStr,
766 SDNode OpNode, bit Commutable = 0> {
768 // Scalar operation, reg+reg.
769 def SSrr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src1, FR32:$src2),
770 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
771 [(set FR32:$dst, (OpNode FR32:$src1, FR32:$src2))]> {
772 let isCommutable = Commutable;
775 def SDrr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src1, FR64:$src2),
776 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
777 [(set FR64:$dst, (OpNode FR64:$src1, FR64:$src2))]> {
778 let isCommutable = Commutable;
781 // Scalar operation, reg+mem.
782 def SSrm : SSI<opc, MRMSrcMem, (outs FR32:$dst),
783 (ins FR32:$src1, f32mem:$src2),
784 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
785 [(set FR32:$dst, (OpNode FR32:$src1, (load addr:$src2)))]>;
787 def SDrm : SDI<opc, MRMSrcMem, (outs FR64:$dst),
788 (ins FR64:$src1, f64mem:$src2),
789 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
790 [(set FR64:$dst, (OpNode FR64:$src1, (load addr:$src2)))]>;
792 // Vector operation, reg+reg.
793 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst),
794 (ins VR128:$src1, VR128:$src2),
795 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
796 [(set VR128:$dst, (v4f32 (OpNode VR128:$src1, VR128:$src2)))]> {
797 let isCommutable = Commutable;
800 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
801 (ins VR128:$src1, VR128:$src2),
802 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
803 [(set VR128:$dst, (v2f64 (OpNode VR128:$src1, VR128:$src2)))]> {
804 let isCommutable = Commutable;
807 // Vector operation, reg+mem.
808 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst),
809 (ins VR128:$src1, f128mem:$src2),
810 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
811 [(set VR128:$dst, (OpNode VR128:$src1, (memopv4f32 addr:$src2)))]>;
813 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
814 (ins VR128:$src1, f128mem:$src2),
815 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
816 [(set VR128:$dst, (OpNode VR128:$src1, (memopv2f64 addr:$src2)))]>;
818 // Intrinsic operation, reg+reg.
819 def SSrr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst),
820 (ins VR128:$src1, VR128:$src2),
821 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
822 [(set VR128:$dst, (!nameconcat<Intrinsic>("int_x86_sse_",
823 !strconcat(OpcodeStr, "_ss")) VR128:$src1,
825 // int_x86_sse_xxx_ss
826 let isCommutable = Commutable;
829 def SDrr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst),
830 (ins VR128:$src1, VR128:$src2),
831 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
832 [(set VR128:$dst, (!nameconcat<Intrinsic>("int_x86_sse2_",
833 !strconcat(OpcodeStr, "_sd")) VR128:$src1,
835 // int_x86_sse2_xxx_sd
836 let isCommutable = Commutable;
839 // Intrinsic operation, reg+mem.
840 def SSrm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst),
841 (ins VR128:$src1, ssmem:$src2),
842 !strconcat(OpcodeStr, "ss\t{$src2, $dst|$dst, $src2}"),
843 [(set VR128:$dst, (!nameconcat<Intrinsic>("int_x86_sse_",
844 !strconcat(OpcodeStr, "_ss")) VR128:$src1,
845 sse_load_f32:$src2))]>;
846 // int_x86_sse_xxx_ss
848 def SDrm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst),
849 (ins VR128:$src1, sdmem:$src2),
850 !strconcat(OpcodeStr, "sd\t{$src2, $dst|$dst, $src2}"),
851 [(set VR128:$dst, (!nameconcat<Intrinsic>("int_x86_sse2_",
852 !strconcat(OpcodeStr, "_sd")) VR128:$src1,
853 sse_load_f64:$src2))]>;
854 // int_x86_sse2_xxx_sd
856 // Vector intrinsic operation, reg+reg.
857 def PSrr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst),
858 (ins VR128:$src1, VR128:$src2),
859 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
860 [(set VR128:$dst, (!nameconcat<Intrinsic>("int_x86_sse_",
861 !strconcat(OpcodeStr, "_ps")) VR128:$src1,
863 // int_x86_sse_xxx_ps
864 let isCommutable = Commutable;
867 def PDrr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst),
868 (ins VR128:$src1, VR128:$src2),
869 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
870 [(set VR128:$dst, (!nameconcat<Intrinsic>("int_x86_sse2_",
871 !strconcat(OpcodeStr, "_pd")) VR128:$src1,
873 // int_x86_sse2_xxx_pd
874 let isCommutable = Commutable;
877 // Vector intrinsic operation, reg+mem.
878 def PSrm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst),
879 (ins VR128:$src1, f128mem:$src2),
880 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
881 [(set VR128:$dst, (!nameconcat<Intrinsic>("int_x86_sse_",
882 !strconcat(OpcodeStr, "_ps")) VR128:$src1,
883 (memopv4f32 addr:$src2)))]>;
884 // int_x86_sse_xxx_ps
886 def PDrm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst),
887 (ins VR128:$src1, f128mem:$src2),
888 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
889 [(set VR128:$dst, (!nameconcat<Intrinsic>("int_x86_sse2_",
890 !strconcat(OpcodeStr, "_pd")) VR128:$src1,
891 (memopv2f64 addr:$src2)))]>;
892 // int_x86_sse2_xxx_pd
896 defm MAX : sse12_fp_binop_rm<0x5F, "max", X86fmax>;
897 defm MIN : sse12_fp_binop_rm<0x5D, "min", X86fmin>;
899 //===----------------------------------------------------------------------===//
900 // SSE packed FP Instructions
903 let neverHasSideEffects = 1 in
904 def MOVAPSrr : PSI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
905 "movaps\t{$src, $dst|$dst, $src}", []>;
906 let canFoldAsLoad = 1, isReMaterializable = 1 in
907 def MOVAPSrm : PSI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
908 "movaps\t{$src, $dst|$dst, $src}",
909 [(set VR128:$dst, (alignedloadv4f32 addr:$src))]>;
911 def MOVAPSmr : PSI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
912 "movaps\t{$src, $dst|$dst, $src}",
913 [(alignedstore (v4f32 VR128:$src), addr:$dst)]>;
915 let neverHasSideEffects = 1 in
916 def MOVUPSrr : PSI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
917 "movups\t{$src, $dst|$dst, $src}", []>;
918 let canFoldAsLoad = 1, isReMaterializable = 1 in
919 def MOVUPSrm : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
920 "movups\t{$src, $dst|$dst, $src}",
921 [(set VR128:$dst, (loadv4f32 addr:$src))]>;
922 def MOVUPSmr : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
923 "movups\t{$src, $dst|$dst, $src}",
924 [(store (v4f32 VR128:$src), addr:$dst)]>;
926 // Intrinsic forms of MOVUPS load and store
927 let canFoldAsLoad = 1, isReMaterializable = 1 in
928 def MOVUPSrm_Int : PSI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
929 "movups\t{$src, $dst|$dst, $src}",
930 [(set VR128:$dst, (int_x86_sse_loadu_ps addr:$src))]>;
931 def MOVUPSmr_Int : PSI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
932 "movups\t{$src, $dst|$dst, $src}",
933 [(int_x86_sse_storeu_ps addr:$dst, VR128:$src)]>;
935 let Constraints = "$src1 = $dst" in {
936 let AddedComplexity = 20 in {
937 def MOVLPSrm : PSI<0x12, MRMSrcMem,
938 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
939 "movlps\t{$src2, $dst|$dst, $src2}",
942 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))]>;
943 def MOVHPSrm : PSI<0x16, MRMSrcMem,
944 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
945 "movhps\t{$src2, $dst|$dst, $src2}",
947 (movlhps VR128:$src1,
948 (bc_v4f32 (v2f64 (scalar_to_vector (loadf64 addr:$src2))))))]>;
950 } // Constraints = "$src1 = $dst"
953 def : Pat<(movlhps VR128:$src1, (bc_v4i32 (v2i64 (X86vzload addr:$src2)))),
954 (MOVHPSrm (v4i32 VR128:$src1), addr:$src2)>;
956 def MOVLPSmr : PSI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
957 "movlps\t{$src, $dst|$dst, $src}",
958 [(store (f64 (vector_extract (bc_v2f64 (v4f32 VR128:$src)),
959 (iPTR 0))), addr:$dst)]>;
961 // v2f64 extract element 1 is always custom lowered to unpack high to low
962 // and extract element 0 so the non-store version isn't too horrible.
963 def MOVHPSmr : PSI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
964 "movhps\t{$src, $dst|$dst, $src}",
965 [(store (f64 (vector_extract
966 (unpckh (bc_v2f64 (v4f32 VR128:$src)),
967 (undef)), (iPTR 0))), addr:$dst)]>;
969 let Constraints = "$src1 = $dst" in {
970 let AddedComplexity = 20 in {
971 def MOVLHPSrr : PSI<0x16, MRMSrcReg, (outs VR128:$dst),
972 (ins VR128:$src1, VR128:$src2),
973 "movlhps\t{$src2, $dst|$dst, $src2}",
975 (v4f32 (movlhps VR128:$src1, VR128:$src2)))]>;
977 def MOVHLPSrr : PSI<0x12, MRMSrcReg, (outs VR128:$dst),
978 (ins VR128:$src1, VR128:$src2),
979 "movhlps\t{$src2, $dst|$dst, $src2}",
981 (v4f32 (movhlps VR128:$src1, VR128:$src2)))]>;
983 } // Constraints = "$src1 = $dst"
985 let AddedComplexity = 20 in {
986 def : Pat<(v4f32 (movddup VR128:$src, (undef))),
987 (MOVLHPSrr (v4f32 VR128:$src), (v4f32 VR128:$src))>;
988 def : Pat<(v2i64 (movddup VR128:$src, (undef))),
989 (MOVLHPSrr (v2i64 VR128:$src), (v2i64 VR128:$src))>;
996 /// sse1_fp_unop_rm - SSE1 unops come in both scalar and vector forms.
998 /// In addition, we also have a special variant of the scalar form here to
999 /// represent the associated intrinsic operation. This form is unlike the
1000 /// plain scalar form, in that it takes an entire vector (instead of a
1001 /// scalar) and leaves the top elements undefined.
1003 /// And, we have a special variant form for a full-vector intrinsic form.
1005 /// These four forms can each have a reg or a mem operand, so there are a
1006 /// total of eight "instructions".
1008 multiclass sse1_fp_unop_rm<bits<8> opc, string OpcodeStr,
1012 bit Commutable = 0> {
1013 // Scalar operation, reg.
1014 def SSr : SSI<opc, MRMSrcReg, (outs FR32:$dst), (ins FR32:$src),
1015 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
1016 [(set FR32:$dst, (OpNode FR32:$src))]> {
1017 let isCommutable = Commutable;
1020 // Scalar operation, mem.
1021 def SSm : I<opc, MRMSrcMem, (outs FR32:$dst), (ins f32mem:$src),
1022 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
1023 [(set FR32:$dst, (OpNode (load addr:$src)))]>, XS,
1024 Requires<[HasSSE1, OptForSize]>;
1026 // Vector operation, reg.
1027 def PSr : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1028 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1029 [(set VR128:$dst, (v4f32 (OpNode VR128:$src)))]> {
1030 let isCommutable = Commutable;
1033 // Vector operation, mem.
1034 def PSm : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1035 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1036 [(set VR128:$dst, (OpNode (memopv4f32 addr:$src)))]>;
1038 // Intrinsic operation, reg.
1039 def SSr_Int : SSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1040 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
1041 [(set VR128:$dst, (F32Int VR128:$src))]> {
1042 let isCommutable = Commutable;
1045 // Intrinsic operation, mem.
1046 def SSm_Int : SSI<opc, MRMSrcMem, (outs VR128:$dst), (ins ssmem:$src),
1047 !strconcat(OpcodeStr, "ss\t{$src, $dst|$dst, $src}"),
1048 [(set VR128:$dst, (F32Int sse_load_f32:$src))]>;
1050 // Vector intrinsic operation, reg
1051 def PSr_Int : PSI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1052 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1053 [(set VR128:$dst, (V4F32Int VR128:$src))]> {
1054 let isCommutable = Commutable;
1057 // Vector intrinsic operation, mem
1058 def PSm_Int : PSI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1059 !strconcat(OpcodeStr, "ps\t{$src, $dst|$dst, $src}"),
1060 [(set VR128:$dst, (V4F32Int (memopv4f32 addr:$src)))]>;
1064 defm SQRT : sse1_fp_unop_rm<0x51, "sqrt", fsqrt,
1065 int_x86_sse_sqrt_ss, int_x86_sse_sqrt_ps>;
1067 // Reciprocal approximations. Note that these typically require refinement
1068 // in order to obtain suitable precision.
1069 defm RSQRT : sse1_fp_unop_rm<0x52, "rsqrt", X86frsqrt,
1070 int_x86_sse_rsqrt_ss, int_x86_sse_rsqrt_ps>;
1071 defm RCP : sse1_fp_unop_rm<0x53, "rcp", X86frcp,
1072 int_x86_sse_rcp_ss, int_x86_sse_rcp_ps>;
1074 /// sse12_fp_pack_logical - SSE 1 & 2 packed FP logical ops
1076 multiclass sse12_fp_pack_logical<bits<8> opc, string OpcodeStr,
1077 SDNode OpNode, int HasPat = 0,
1079 list<list<dag>> Pattern = []> {
1080 def PSrr : PSI<opc, MRMSrcReg, (outs VR128:$dst),
1081 (ins VR128:$src1, VR128:$src2),
1082 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
1083 !if(HasPat, Pattern[0],
1084 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1,
1086 { let isCommutable = Commutable; }
1088 def PDrr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1089 (ins VR128:$src1, VR128:$src2),
1090 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1091 !if(HasPat, Pattern[1],
1092 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
1093 (bc_v2i64 (v2f64 VR128:$src2))))])>
1094 { let isCommutable = Commutable; }
1096 def PSrm : PSI<opc, MRMSrcMem, (outs VR128:$dst),
1097 (ins VR128:$src1, f128mem:$src2),
1098 !strconcat(OpcodeStr, "ps\t{$src2, $dst|$dst, $src2}"),
1099 !if(HasPat, Pattern[2],
1100 [(set VR128:$dst, (OpNode (bc_v2i64 (v4f32 VR128:$src1)),
1101 (memopv2i64 addr:$src2)))])>;
1103 def PDrm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1104 (ins VR128:$src1, f128mem:$src2),
1105 !strconcat(OpcodeStr, "pd\t{$src2, $dst|$dst, $src2}"),
1106 !if(HasPat, Pattern[3],
1107 [(set VR128:$dst, (OpNode (bc_v2i64 (v2f64 VR128:$src1)),
1108 (memopv2i64 addr:$src2)))])>;
1112 let Constraints = "$src1 = $dst" in {
1113 defm AND : sse12_fp_pack_logical<0x54, "and", and>;
1114 defm OR : sse12_fp_pack_logical<0x56, "or", or>;
1115 defm XOR : sse12_fp_pack_logical<0x57, "xor", xor>;
1116 defm ANDN : sse12_fp_pack_logical<0x55, "andn", undef /* dummy */, 1, 0, [
1118 [(set VR128:$dst, (v2i64 (and (xor VR128:$src1,
1119 (bc_v2i64 (v4i32 immAllOnesV))),
1122 [(set VR128:$dst, (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1123 (bc_v2i64 (v2f64 VR128:$src2))))],
1125 [(set VR128:$dst, (v2i64 (and (xor (bc_v2i64 (v4f32 VR128:$src1)),
1126 (bc_v2i64 (v4i32 immAllOnesV))),
1127 (memopv2i64 addr:$src2))))],
1129 [(set VR128:$dst, (and (vnot (bc_v2i64 (v2f64 VR128:$src1))),
1130 (memopv2i64 addr:$src2)))]]>;
1133 let Constraints = "$src1 = $dst" in {
1134 def CMPPSrri : PSIi8<0xC2, MRMSrcReg,
1135 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1136 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
1137 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
1138 VR128:$src, imm:$cc))]>;
1139 def CMPPSrmi : PSIi8<0xC2, MRMSrcMem,
1140 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
1141 "cmp${cc}ps\t{$src, $dst|$dst, $src}",
1142 [(set VR128:$dst, (int_x86_sse_cmp_ps VR128:$src1,
1143 (memop addr:$src), imm:$cc))]>;
1145 // Accept explicit immediate argument form instead of comparison code.
1146 let isAsmParserOnly = 1 in {
1147 def CMPPSrri_alt : PSIi8<0xC2, MRMSrcReg,
1148 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, i8imm:$src2),
1149 "cmpps\t{$src2, $src, $dst|$dst, $src, $src}", []>;
1150 def CMPPSrmi_alt : PSIi8<0xC2, MRMSrcMem,
1151 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, i8imm:$src2),
1152 "cmpps\t{$src2, $src, $dst|$dst, $src, $src}", []>;
1155 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), VR128:$src2, imm:$cc)),
1156 (CMPPSrri (v4f32 VR128:$src1), (v4f32 VR128:$src2), imm:$cc)>;
1157 def : Pat<(v4i32 (X86cmpps (v4f32 VR128:$src1), (memop addr:$src2), imm:$cc)),
1158 (CMPPSrmi (v4f32 VR128:$src1), addr:$src2, imm:$cc)>;
1160 // Shuffle and unpack instructions
1161 let Constraints = "$src1 = $dst" in {
1162 let isConvertibleToThreeAddress = 1 in // Convert to pshufd
1163 def SHUFPSrri : PSIi8<0xC6, MRMSrcReg,
1164 (outs VR128:$dst), (ins VR128:$src1,
1165 VR128:$src2, i8imm:$src3),
1166 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1168 (v4f32 (shufp:$src3 VR128:$src1, VR128:$src2)))]>;
1169 def SHUFPSrmi : PSIi8<0xC6, MRMSrcMem,
1170 (outs VR128:$dst), (ins VR128:$src1,
1171 f128mem:$src2, i8imm:$src3),
1172 "shufps\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1175 VR128:$src1, (memopv4f32 addr:$src2))))]>;
1177 let AddedComplexity = 10 in {
1178 def UNPCKHPSrr : PSI<0x15, MRMSrcReg,
1179 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1180 "unpckhps\t{$src2, $dst|$dst, $src2}",
1182 (v4f32 (unpckh VR128:$src1, VR128:$src2)))]>;
1183 def UNPCKHPSrm : PSI<0x15, MRMSrcMem,
1184 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1185 "unpckhps\t{$src2, $dst|$dst, $src2}",
1187 (v4f32 (unpckh VR128:$src1,
1188 (memopv4f32 addr:$src2))))]>;
1190 def UNPCKLPSrr : PSI<0x14, MRMSrcReg,
1191 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1192 "unpcklps\t{$src2, $dst|$dst, $src2}",
1194 (v4f32 (unpckl VR128:$src1, VR128:$src2)))]>;
1195 def UNPCKLPSrm : PSI<0x14, MRMSrcMem,
1196 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1197 "unpcklps\t{$src2, $dst|$dst, $src2}",
1199 (unpckl VR128:$src1, (memopv4f32 addr:$src2)))]>;
1200 } // AddedComplexity
1201 } // Constraints = "$src1 = $dst"
1204 def MOVMSKPSrr : PSI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1205 "movmskps\t{$src, $dst|$dst, $src}",
1206 [(set GR32:$dst, (int_x86_sse_movmsk_ps VR128:$src))]>;
1207 def MOVMSKPDrr : PDI<0x50, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1208 "movmskpd\t{$src, $dst|$dst, $src}",
1209 [(set GR32:$dst, (int_x86_sse2_movmsk_pd VR128:$src))]>;
1211 // Prefetch intrinsic.
1212 def PREFETCHT0 : PSI<0x18, MRM1m, (outs), (ins i8mem:$src),
1213 "prefetcht0\t$src", [(prefetch addr:$src, imm, (i32 3))]>;
1214 def PREFETCHT1 : PSI<0x18, MRM2m, (outs), (ins i8mem:$src),
1215 "prefetcht1\t$src", [(prefetch addr:$src, imm, (i32 2))]>;
1216 def PREFETCHT2 : PSI<0x18, MRM3m, (outs), (ins i8mem:$src),
1217 "prefetcht2\t$src", [(prefetch addr:$src, imm, (i32 1))]>;
1218 def PREFETCHNTA : PSI<0x18, MRM0m, (outs), (ins i8mem:$src),
1219 "prefetchnta\t$src", [(prefetch addr:$src, imm, (i32 0))]>;
1221 // Non-temporal stores
1222 def MOVNTPSmr_Int : PSI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1223 "movntps\t{$src, $dst|$dst, $src}",
1224 [(int_x86_sse_movnt_ps addr:$dst, VR128:$src)]>;
1226 let AddedComplexity = 400 in { // Prefer non-temporal versions
1227 def MOVNTPSmr : PSI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1228 "movntps\t{$src, $dst|$dst, $src}",
1229 [(alignednontemporalstore (v4f32 VR128:$src), addr:$dst)]>;
1231 def MOVNTDQ_64mr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1232 "movntdq\t{$src, $dst|$dst, $src}",
1233 [(alignednontemporalstore (v2f64 VR128:$src), addr:$dst)]>;
1235 def MOVNTImr : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
1236 "movnti\t{$src, $dst|$dst, $src}",
1237 [(nontemporalstore (i32 GR32:$src), addr:$dst)]>,
1238 TB, Requires<[HasSSE2]>;
1240 def MOVNTI_64mr : RI<0xC3, MRMDestMem, (outs), (ins i64mem:$dst, GR64:$src),
1241 "movnti\t{$src, $dst|$dst, $src}",
1242 [(nontemporalstore (i64 GR64:$src), addr:$dst)]>,
1243 TB, Requires<[HasSSE2]>;
1246 // Load, store, and memory fence
1247 def SFENCE : I<0xAE, MRM_F8, (outs), (ins), "sfence", [(int_x86_sse_sfence)]>,
1248 TB, Requires<[HasSSE1]>;
1251 def LDMXCSR : PSI<0xAE, MRM2m, (outs), (ins i32mem:$src),
1252 "ldmxcsr\t$src", [(int_x86_sse_ldmxcsr addr:$src)]>;
1253 def STMXCSR : PSI<0xAE, MRM3m, (outs), (ins i32mem:$dst),
1254 "stmxcsr\t$dst", [(int_x86_sse_stmxcsr addr:$dst)]>;
1256 // Alias instructions that map zero vector to pxor / xorp* for sse.
1257 // We set canFoldAsLoad because this can be converted to a constant-pool
1258 // load of an all-zeros value if folding it would be beneficial.
1259 // FIXME: Change encoding to pseudo!
1260 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
1261 isCodeGenOnly = 1 in {
1262 def V_SET0PS : PSI<0x57, MRMInitReg, (outs VR128:$dst), (ins), "",
1263 [(set VR128:$dst, (v4f32 immAllZerosV))]>;
1264 def V_SET0PD : PDI<0x57, MRMInitReg, (outs VR128:$dst), (ins), "",
1265 [(set VR128:$dst, (v2f64 immAllZerosV))]>;
1266 let ExeDomain = SSEPackedInt in
1267 def V_SET0PI : PDI<0xEF, MRMInitReg, (outs VR128:$dst), (ins), "",
1268 [(set VR128:$dst, (v4i32 immAllZerosV))]>;
1271 def : Pat<(v2i64 immAllZerosV), (V_SET0PI)>;
1272 def : Pat<(v8i16 immAllZerosV), (V_SET0PI)>;
1273 def : Pat<(v16i8 immAllZerosV), (V_SET0PI)>;
1275 def : Pat<(f32 (vector_extract (v4f32 VR128:$src), (iPTR 0))),
1276 (f32 (EXTRACT_SUBREG (v4f32 VR128:$src), sub_ss))>;
1278 //===---------------------------------------------------------------------===//
1279 // SSE2 Instructions
1280 //===---------------------------------------------------------------------===//
1282 // Move Instructions. Register-to-register movsd is not used for FR64
1283 // register copies because it's a partial register update; FsMOVAPDrr is
1284 // used instead. Register-to-register movsd is not modeled as an INSERT_SUBREG
1285 // because INSERT_SUBREG requires that the insert be implementable in terms of
1286 // a copy, and just mentioned, we don't use movsd for copies.
1287 let Constraints = "$src1 = $dst" in
1288 def MOVSDrr : SDI<0x10, MRMSrcReg,
1289 (outs VR128:$dst), (ins VR128:$src1, FR64:$src2),
1290 "movsd\t{$src2, $dst|$dst, $src2}",
1291 [(set (v2f64 VR128:$dst),
1292 (movl VR128:$src1, (scalar_to_vector FR64:$src2)))]>;
1294 // Extract the low 64-bit value from one vector and insert it into another.
1295 let AddedComplexity = 15 in
1296 def : Pat<(v2f64 (movl VR128:$src1, VR128:$src2)),
1297 (MOVSDrr (v2f64 VR128:$src1),
1298 (EXTRACT_SUBREG (v2f64 VR128:$src2), sub_sd))>;
1300 // Implicitly promote a 64-bit scalar to a vector.
1301 def : Pat<(v2f64 (scalar_to_vector FR64:$src)),
1302 (INSERT_SUBREG (v2f64 (IMPLICIT_DEF)), FR64:$src, sub_sd)>;
1304 // Loading from memory automatically zeroing upper bits.
1305 let canFoldAsLoad = 1, isReMaterializable = 1, AddedComplexity = 20 in
1306 def MOVSDrm : SDI<0x10, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1307 "movsd\t{$src, $dst|$dst, $src}",
1308 [(set FR64:$dst, (loadf64 addr:$src))]>;
1310 // MOVSDrm zeros the high parts of the register; represent this
1311 // with SUBREG_TO_REG.
1312 let AddedComplexity = 20 in {
1313 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector (loadf64 addr:$src))))),
1314 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
1315 def : Pat<(v2f64 (scalar_to_vector (loadf64 addr:$src))),
1316 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
1317 def : Pat<(v2f64 (X86vzmovl (loadv2f64 addr:$src))),
1318 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
1319 def : Pat<(v2f64 (X86vzmovl (bc_v2f64 (loadv4f32 addr:$src)))),
1320 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
1321 def : Pat<(v2f64 (X86vzload addr:$src)),
1322 (SUBREG_TO_REG (i64 0), (MOVSDrm addr:$src), sub_sd)>;
1325 // Store scalar value to memory.
1326 def MOVSDmr : SDI<0x11, MRMDestMem, (outs), (ins f64mem:$dst, FR64:$src),
1327 "movsd\t{$src, $dst|$dst, $src}",
1328 [(store FR64:$src, addr:$dst)]>;
1330 // Extract and store.
1331 def : Pat<(store (f64 (vector_extract (v2f64 VR128:$src), (iPTR 0))),
1334 (EXTRACT_SUBREG (v2f64 VR128:$src), sub_sd))>;
1336 // Conversion instructions
1337 def CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins FR64:$src),
1338 "cvttsd2si\t{$src, $dst|$dst, $src}",
1339 [(set GR32:$dst, (fp_to_sint FR64:$src))]>;
1340 def CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f64mem:$src),
1341 "cvttsd2si\t{$src, $dst|$dst, $src}",
1342 [(set GR32:$dst, (fp_to_sint (loadf64 addr:$src)))]>;
1343 def CVTSD2SSrr : SDI<0x5A, MRMSrcReg, (outs FR32:$dst), (ins FR64:$src),
1344 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1345 [(set FR32:$dst, (fround FR64:$src))]>;
1346 def CVTSD2SSrm : I<0x5A, MRMSrcMem, (outs FR32:$dst), (ins f64mem:$src),
1347 "cvtsd2ss\t{$src, $dst|$dst, $src}",
1348 [(set FR32:$dst, (fround (loadf64 addr:$src)))]>, XD,
1349 Requires<[HasSSE2, OptForSize]>;
1350 def CVTSI2SDrr : SDI<0x2A, MRMSrcReg, (outs FR64:$dst), (ins GR32:$src),
1351 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1352 [(set FR64:$dst, (sint_to_fp GR32:$src))]>;
1353 def CVTSI2SDrm : SDI<0x2A, MRMSrcMem, (outs FR64:$dst), (ins i32mem:$src),
1354 "cvtsi2sd\t{$src, $dst|$dst, $src}",
1355 [(set FR64:$dst, (sint_to_fp (loadi32 addr:$src)))]>;
1357 def CVTPD2DQrm : S3DI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1358 "cvtpd2dq\t{$src, $dst|$dst, $src}", []>;
1359 def CVTPD2DQrr : S3DI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1360 "cvtpd2dq\t{$src, $dst|$dst, $src}", []>;
1361 def CVTDQ2PDrm : S3SI<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1362 "cvtdq2pd\t{$src, $dst|$dst, $src}", []>;
1363 def CVTDQ2PDrr : S3SI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1364 "cvtdq2pd\t{$src, $dst|$dst, $src}", []>;
1365 def CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1366 "cvtps2dq\t{$src, $dst|$dst, $src}", []>;
1367 def CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1368 "cvtps2dq\t{$src, $dst|$dst, $src}", []>;
1369 def CVTDQ2PSrr : PSI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1370 "cvtdq2ps\t{$src, $dst|$dst, $src}", []>;
1371 def CVTDQ2PSrm : PSI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1372 "cvtdq2ps\t{$src, $dst|$dst, $src}", []>;
1373 def COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1374 "comisd\t{$src2, $src1|$src1, $src2}", []>;
1375 def COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1376 "comisd\t{$src2, $src1|$src1, $src2}", []>;
1378 // SSE2 instructions with XS prefix
1379 def CVTSS2SDrr : I<0x5A, MRMSrcReg, (outs FR64:$dst), (ins FR32:$src),
1380 "cvtss2sd\t{$src, $dst|$dst, $src}",
1381 [(set FR64:$dst, (fextend FR32:$src))]>, XS,
1382 Requires<[HasSSE2]>;
1383 def CVTSS2SDrm : I<0x5A, MRMSrcMem, (outs FR64:$dst), (ins f32mem:$src),
1384 "cvtss2sd\t{$src, $dst|$dst, $src}",
1385 [(set FR64:$dst, (extloadf32 addr:$src))]>, XS,
1386 Requires<[HasSSE2, OptForSize]>;
1388 def : Pat<(extloadf32 addr:$src),
1389 (CVTSS2SDrr (MOVSSrm addr:$src))>,
1390 Requires<[HasSSE2, OptForSpeed]>;
1392 // Match intrinsics which expect XMM operand(s).
1393 def Int_CVTSD2SIrr : SDI<0x2D, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1394 "cvtsd2si\t{$src, $dst|$dst, $src}",
1395 [(set GR32:$dst, (int_x86_sse2_cvtsd2si VR128:$src))]>;
1396 def Int_CVTSD2SIrm : SDI<0x2D, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1397 "cvtsd2si\t{$src, $dst|$dst, $src}",
1398 [(set GR32:$dst, (int_x86_sse2_cvtsd2si
1399 (load addr:$src)))]>;
1401 // Match intrinsics which expect MM and XMM operand(s).
1402 def Int_CVTPD2PIrr : PDI<0x2D, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1403 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1404 [(set VR64:$dst, (int_x86_sse_cvtpd2pi VR128:$src))]>;
1405 def Int_CVTPD2PIrm : PDI<0x2D, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1406 "cvtpd2pi\t{$src, $dst|$dst, $src}",
1407 [(set VR64:$dst, (int_x86_sse_cvtpd2pi
1408 (memop addr:$src)))]>;
1409 def Int_CVTTPD2PIrr: PDI<0x2C, MRMSrcReg, (outs VR64:$dst), (ins VR128:$src),
1410 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1411 [(set VR64:$dst, (int_x86_sse_cvttpd2pi VR128:$src))]>;
1412 def Int_CVTTPD2PIrm: PDI<0x2C, MRMSrcMem, (outs VR64:$dst), (ins f128mem:$src),
1413 "cvttpd2pi\t{$src, $dst|$dst, $src}",
1414 [(set VR64:$dst, (int_x86_sse_cvttpd2pi
1415 (memop addr:$src)))]>;
1416 def Int_CVTPI2PDrr : PDI<0x2A, MRMSrcReg, (outs VR128:$dst), (ins VR64:$src),
1417 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1418 [(set VR128:$dst, (int_x86_sse_cvtpi2pd VR64:$src))]>;
1419 def Int_CVTPI2PDrm : PDI<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1420 "cvtpi2pd\t{$src, $dst|$dst, $src}",
1421 [(set VR128:$dst, (int_x86_sse_cvtpi2pd
1422 (load addr:$src)))]>;
1424 // Aliases for intrinsics
1425 def Int_CVTTSD2SIrr : SDI<0x2C, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
1426 "cvttsd2si\t{$src, $dst|$dst, $src}",
1428 (int_x86_sse2_cvttsd2si VR128:$src))]>;
1429 def Int_CVTTSD2SIrm : SDI<0x2C, MRMSrcMem, (outs GR32:$dst), (ins f128mem:$src),
1430 "cvttsd2si\t{$src, $dst|$dst, $src}",
1431 [(set GR32:$dst, (int_x86_sse2_cvttsd2si
1432 (load addr:$src)))]>;
1434 // Comparison instructions
1435 let Constraints = "$src1 = $dst", neverHasSideEffects = 1 in {
1436 def CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1437 (outs FR64:$dst), (ins FR64:$src1, FR64:$src, SSECC:$cc),
1438 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1440 def CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1441 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src, SSECC:$cc),
1442 "cmp${cc}sd\t{$src, $dst|$dst, $src}", []>;
1444 // Accept explicit immediate argument form instead of comparison code.
1445 let isAsmParserOnly = 1 in {
1446 def CMPSDrr_alt : SDIi8<0xC2, MRMSrcReg,
1447 (outs FR64:$dst), (ins FR64:$src1, FR64:$src, i8imm:$src2),
1448 "cmpsd\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
1450 def CMPSDrm_alt : SDIi8<0xC2, MRMSrcMem,
1451 (outs FR64:$dst), (ins FR64:$src1, f64mem:$src, i8imm:$src2),
1452 "cmpsd\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
1456 let Defs = [EFLAGS] in {
1457 def UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins FR64:$src1, FR64:$src2),
1458 "ucomisd\t{$src2, $src1|$src1, $src2}",
1459 [(set EFLAGS, (X86cmp FR64:$src1, FR64:$src2))]>;
1460 def UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs), (ins FR64:$src1, f64mem:$src2),
1461 "ucomisd\t{$src2, $src1|$src1, $src2}",
1462 [(set EFLAGS, (X86cmp FR64:$src1, (loadf64 addr:$src2)))]>;
1463 } // Defs = [EFLAGS]
1465 // Aliases to match intrinsics which expect XMM operand(s).
1466 let Constraints = "$src1 = $dst" in {
1467 def Int_CMPSDrr : SDIi8<0xC2, MRMSrcReg,
1469 (ins VR128:$src1, VR128:$src, SSECC:$cc),
1470 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1471 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1472 VR128:$src, imm:$cc))]>;
1473 def Int_CMPSDrm : SDIi8<0xC2, MRMSrcMem,
1475 (ins VR128:$src1, f64mem:$src, SSECC:$cc),
1476 "cmp${cc}sd\t{$src, $dst|$dst, $src}",
1477 [(set VR128:$dst, (int_x86_sse2_cmp_sd VR128:$src1,
1478 (load addr:$src), imm:$cc))]>;
1481 let Defs = [EFLAGS] in {
1482 def Int_UCOMISDrr: PDI<0x2E, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1483 "ucomisd\t{$src2, $src1|$src1, $src2}",
1484 [(set EFLAGS, (X86ucomi (v2f64 VR128:$src1),
1486 def Int_UCOMISDrm: PDI<0x2E, MRMSrcMem, (outs),(ins VR128:$src1, f128mem:$src2),
1487 "ucomisd\t{$src2, $src1|$src1, $src2}",
1488 [(set EFLAGS, (X86ucomi (v2f64 VR128:$src1),
1489 (load addr:$src2)))]>;
1491 def Int_COMISDrr: PDI<0x2F, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
1492 "comisd\t{$src2, $src1|$src1, $src2}",
1493 [(set EFLAGS, (X86comi (v2f64 VR128:$src1),
1495 def Int_COMISDrm: PDI<0x2F, MRMSrcMem, (outs), (ins VR128:$src1, f128mem:$src2),
1496 "comisd\t{$src2, $src1|$src1, $src2}",
1497 [(set EFLAGS, (X86comi (v2f64 VR128:$src1),
1498 (load addr:$src2)))]>;
1499 } // Defs = [EFLAGS]
1501 // Aliases of packed SSE2 instructions for scalar use. These all have names
1502 // that start with 'Fs'.
1504 // Alias instructions that map fld0 to pxor for sse.
1505 let isReMaterializable = 1, isAsCheapAsAMove = 1, isCodeGenOnly = 1,
1506 canFoldAsLoad = 1 in
1507 def FsFLD0SD : I<0xEF, MRMInitReg, (outs FR64:$dst), (ins), "",
1508 [(set FR64:$dst, fpimm0)]>,
1509 Requires<[HasSSE2]>, TB, OpSize;
1511 // Alias instruction to do FR64 reg-to-reg copy using movapd. Upper bits are
1513 let neverHasSideEffects = 1 in
1514 def FsMOVAPDrr : PDI<0x28, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1515 "movapd\t{$src, $dst|$dst, $src}", []>;
1517 // Alias instruction to load FR64 from f128mem using movapd. Upper bits are
1519 let canFoldAsLoad = 1, isReMaterializable = 1 in
1520 def FsMOVAPDrm : PDI<0x28, MRMSrcMem, (outs FR64:$dst), (ins f128mem:$src),
1521 "movapd\t{$src, $dst|$dst, $src}",
1522 [(set FR64:$dst, (alignedloadfsf64 addr:$src))]>;
1524 //===---------------------------------------------------------------------===//
1525 // SSE packed FP Instructions
1527 // Move Instructions
1528 let neverHasSideEffects = 1 in
1529 def MOVAPDrr : PDI<0x28, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1530 "movapd\t{$src, $dst|$dst, $src}", []>;
1531 let canFoldAsLoad = 1, isReMaterializable = 1 in
1532 def MOVAPDrm : PDI<0x28, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1533 "movapd\t{$src, $dst|$dst, $src}",
1534 [(set VR128:$dst, (alignedloadv2f64 addr:$src))]>;
1536 def MOVAPDmr : PDI<0x29, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1537 "movapd\t{$src, $dst|$dst, $src}",
1538 [(alignedstore (v2f64 VR128:$src), addr:$dst)]>;
1540 let neverHasSideEffects = 1 in
1541 def MOVUPDrr : PDI<0x10, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1542 "movupd\t{$src, $dst|$dst, $src}", []>;
1543 let canFoldAsLoad = 1 in
1544 def MOVUPDrm : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1545 "movupd\t{$src, $dst|$dst, $src}",
1546 [(set VR128:$dst, (loadv2f64 addr:$src))]>;
1547 def MOVUPDmr : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1548 "movupd\t{$src, $dst|$dst, $src}",
1549 [(store (v2f64 VR128:$src), addr:$dst)]>;
1551 // Intrinsic forms of MOVUPD load and store
1552 def MOVUPDrm_Int : PDI<0x10, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1553 "movupd\t{$src, $dst|$dst, $src}",
1554 [(set VR128:$dst, (int_x86_sse2_loadu_pd addr:$src))]>;
1555 def MOVUPDmr_Int : PDI<0x11, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
1556 "movupd\t{$src, $dst|$dst, $src}",
1557 [(int_x86_sse2_storeu_pd addr:$dst, VR128:$src)]>;
1559 let Constraints = "$src1 = $dst" in {
1560 let AddedComplexity = 20 in {
1561 def MOVLPDrm : PDI<0x12, MRMSrcMem,
1562 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1563 "movlpd\t{$src2, $dst|$dst, $src2}",
1565 (v2f64 (movlp VR128:$src1,
1566 (scalar_to_vector (loadf64 addr:$src2)))))]>;
1567 def MOVHPDrm : PDI<0x16, MRMSrcMem,
1568 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1569 "movhpd\t{$src2, $dst|$dst, $src2}",
1571 (v2f64 (movlhps VR128:$src1,
1572 (scalar_to_vector (loadf64 addr:$src2)))))]>;
1573 } // AddedComplexity
1574 } // Constraints = "$src1 = $dst"
1576 def MOVLPDmr : PDI<0x13, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1577 "movlpd\t{$src, $dst|$dst, $src}",
1578 [(store (f64 (vector_extract (v2f64 VR128:$src),
1579 (iPTR 0))), addr:$dst)]>;
1581 // v2f64 extract element 1 is always custom lowered to unpack high to low
1582 // and extract element 0 so the non-store version isn't too horrible.
1583 def MOVHPDmr : PDI<0x17, MRMDestMem, (outs), (ins f64mem:$dst, VR128:$src),
1584 "movhpd\t{$src, $dst|$dst, $src}",
1585 [(store (f64 (vector_extract
1586 (v2f64 (unpckh VR128:$src, (undef))),
1587 (iPTR 0))), addr:$dst)]>;
1589 // SSE2 instructions without OpSize prefix
1590 def Int_CVTDQ2PSrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1591 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1592 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps VR128:$src))]>,
1593 TB, Requires<[HasSSE2]>;
1594 def Int_CVTDQ2PSrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1595 "cvtdq2ps\t{$src, $dst|$dst, $src}",
1596 [(set VR128:$dst, (int_x86_sse2_cvtdq2ps
1597 (bitconvert (memopv2i64 addr:$src))))]>,
1598 TB, Requires<[HasSSE2]>;
1600 // SSE2 instructions with XS prefix
1601 def Int_CVTDQ2PDrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1602 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1603 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd VR128:$src))]>,
1604 XS, Requires<[HasSSE2]>;
1605 def Int_CVTDQ2PDrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
1606 "cvtdq2pd\t{$src, $dst|$dst, $src}",
1607 [(set VR128:$dst, (int_x86_sse2_cvtdq2pd
1608 (bitconvert (memopv2i64 addr:$src))))]>,
1609 XS, Requires<[HasSSE2]>;
1611 def Int_CVTPS2DQrr : PDI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1612 "cvtps2dq\t{$src, $dst|$dst, $src}",
1613 [(set VR128:$dst, (int_x86_sse2_cvtps2dq VR128:$src))]>;
1614 def Int_CVTPS2DQrm : PDI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1615 "cvtps2dq\t{$src, $dst|$dst, $src}",
1616 [(set VR128:$dst, (int_x86_sse2_cvtps2dq
1617 (memop addr:$src)))]>;
1618 // SSE2 packed instructions with XS prefix
1619 def CVTTPS2DQrr : SSI<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1620 "cvttps2dq\t{$src, $dst|$dst, $src}", []>;
1621 def CVTTPS2DQrm : SSI<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1622 "cvttps2dq\t{$src, $dst|$dst, $src}", []>;
1624 def Int_CVTTPS2DQrr : I<0x5B, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1625 "cvttps2dq\t{$src, $dst|$dst, $src}",
1627 (int_x86_sse2_cvttps2dq VR128:$src))]>,
1628 XS, Requires<[HasSSE2]>;
1629 def Int_CVTTPS2DQrm : I<0x5B, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1630 "cvttps2dq\t{$src, $dst|$dst, $src}",
1631 [(set VR128:$dst, (int_x86_sse2_cvttps2dq
1632 (memop addr:$src)))]>,
1633 XS, Requires<[HasSSE2]>;
1635 // SSE2 packed instructions with XD prefix
1636 def Int_CVTPD2DQrr : I<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1637 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1638 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq VR128:$src))]>,
1639 XD, Requires<[HasSSE2]>;
1640 def Int_CVTPD2DQrm : I<0xE6, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1641 "cvtpd2dq\t{$src, $dst|$dst, $src}",
1642 [(set VR128:$dst, (int_x86_sse2_cvtpd2dq
1643 (memop addr:$src)))]>,
1644 XD, Requires<[HasSSE2]>;
1646 def Int_CVTTPD2DQrr : PDI<0xE6, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1647 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1648 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq VR128:$src))]>;
1649 def Int_CVTTPD2DQrm : PDI<0xE6, MRMSrcMem, (outs VR128:$dst),(ins f128mem:$src),
1650 "cvttpd2dq\t{$src, $dst|$dst, $src}",
1651 [(set VR128:$dst, (int_x86_sse2_cvttpd2dq
1652 (memop addr:$src)))]>;
1654 // SSE2 instructions without OpSize prefix
1655 def CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1656 "cvtps2pd\t{$src, $dst|$dst, $src}", []>, TB;
1657 def CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
1658 "cvtps2pd\t{$src, $dst|$dst, $src}", []>, TB;
1660 def Int_CVTPS2PDrr : I<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1661 "cvtps2pd\t{$src, $dst|$dst, $src}",
1662 [(set VR128:$dst, (int_x86_sse2_cvtps2pd VR128:$src))]>,
1663 TB, Requires<[HasSSE2]>;
1664 def Int_CVTPS2PDrm : I<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
1665 "cvtps2pd\t{$src, $dst|$dst, $src}",
1666 [(set VR128:$dst, (int_x86_sse2_cvtps2pd
1667 (load addr:$src)))]>,
1668 TB, Requires<[HasSSE2]>;
1670 def CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1671 "cvtpd2ps\t{$src, $dst|$dst, $src}", []>;
1672 def CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1673 "cvtpd2ps\t{$src, $dst|$dst, $src}", []>;
1676 def Int_CVTPD2PSrr : PDI<0x5A, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1677 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1678 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps VR128:$src))]>;
1679 def Int_CVTPD2PSrm : PDI<0x5A, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1680 "cvtpd2ps\t{$src, $dst|$dst, $src}",
1681 [(set VR128:$dst, (int_x86_sse2_cvtpd2ps
1682 (memop addr:$src)))]>;
1684 // Match intrinsics which expect XMM operand(s).
1685 // Aliases for intrinsics
1686 let Constraints = "$src1 = $dst" in {
1687 def Int_CVTSI2SDrr: SDI<0x2A, MRMSrcReg,
1688 (outs VR128:$dst), (ins VR128:$src1, GR32:$src2),
1689 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1690 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1692 def Int_CVTSI2SDrm: SDI<0x2A, MRMSrcMem,
1693 (outs VR128:$dst), (ins VR128:$src1, i32mem:$src2),
1694 "cvtsi2sd\t{$src2, $dst|$dst, $src2}",
1695 [(set VR128:$dst, (int_x86_sse2_cvtsi2sd VR128:$src1,
1696 (loadi32 addr:$src2)))]>;
1697 def Int_CVTSD2SSrr: SDI<0x5A, MRMSrcReg,
1698 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1699 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1700 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1702 def Int_CVTSD2SSrm: SDI<0x5A, MRMSrcMem,
1703 (outs VR128:$dst), (ins VR128:$src1, f64mem:$src2),
1704 "cvtsd2ss\t{$src2, $dst|$dst, $src2}",
1705 [(set VR128:$dst, (int_x86_sse2_cvtsd2ss VR128:$src1,
1706 (load addr:$src2)))]>;
1707 def Int_CVTSS2SDrr: I<0x5A, MRMSrcReg,
1708 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1709 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1710 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1711 VR128:$src2))]>, XS,
1712 Requires<[HasSSE2]>;
1713 def Int_CVTSS2SDrm: I<0x5A, MRMSrcMem,
1714 (outs VR128:$dst), (ins VR128:$src1, f32mem:$src2),
1715 "cvtss2sd\t{$src2, $dst|$dst, $src2}",
1716 [(set VR128:$dst, (int_x86_sse2_cvtss2sd VR128:$src1,
1717 (load addr:$src2)))]>, XS,
1718 Requires<[HasSSE2]>;
1723 /// sse2_fp_unop_rm - SSE2 unops come in both scalar and vector forms.
1725 /// In addition, we also have a special variant of the scalar form here to
1726 /// represent the associated intrinsic operation. This form is unlike the
1727 /// plain scalar form, in that it takes an entire vector (instead of a
1728 /// scalar) and leaves the top elements undefined.
1730 /// And, we have a special variant form for a full-vector intrinsic form.
1732 /// These four forms can each have a reg or a mem operand, so there are a
1733 /// total of eight "instructions".
1735 multiclass sse2_fp_unop_rm<bits<8> opc, string OpcodeStr,
1739 bit Commutable = 0> {
1740 // Scalar operation, reg.
1741 def SDr : SDI<opc, MRMSrcReg, (outs FR64:$dst), (ins FR64:$src),
1742 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1743 [(set FR64:$dst, (OpNode FR64:$src))]> {
1744 let isCommutable = Commutable;
1747 // Scalar operation, mem.
1748 def SDm : SDI<opc, MRMSrcMem, (outs FR64:$dst), (ins f64mem:$src),
1749 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1750 [(set FR64:$dst, (OpNode (load addr:$src)))]>;
1752 // Vector operation, reg.
1753 def PDr : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1754 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1755 [(set VR128:$dst, (v2f64 (OpNode VR128:$src)))]> {
1756 let isCommutable = Commutable;
1759 // Vector operation, mem.
1760 def PDm : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1761 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1762 [(set VR128:$dst, (OpNode (memopv2f64 addr:$src)))]>;
1764 // Intrinsic operation, reg.
1765 def SDr_Int : SDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1766 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1767 [(set VR128:$dst, (F64Int VR128:$src))]> {
1768 let isCommutable = Commutable;
1771 // Intrinsic operation, mem.
1772 def SDm_Int : SDI<opc, MRMSrcMem, (outs VR128:$dst), (ins sdmem:$src),
1773 !strconcat(OpcodeStr, "sd\t{$src, $dst|$dst, $src}"),
1774 [(set VR128:$dst, (F64Int sse_load_f64:$src))]>;
1776 // Vector intrinsic operation, reg
1777 def PDr_Int : PDI<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1778 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1779 [(set VR128:$dst, (V2F64Int VR128:$src))]> {
1780 let isCommutable = Commutable;
1783 // Vector intrinsic operation, mem
1784 def PDm_Int : PDI<opc, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
1785 !strconcat(OpcodeStr, "pd\t{$src, $dst|$dst, $src}"),
1786 [(set VR128:$dst, (V2F64Int (memopv2f64 addr:$src)))]>;
1790 defm SQRT : sse2_fp_unop_rm<0x51, "sqrt", fsqrt,
1791 int_x86_sse2_sqrt_sd, int_x86_sse2_sqrt_pd>;
1793 // There is no f64 version of the reciprocal approximation instructions.
1795 let Constraints = "$src1 = $dst" in {
1796 def CMPPDrri : PDIi8<0xC2, MRMSrcReg,
1797 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, SSECC:$cc),
1798 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1799 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1800 VR128:$src, imm:$cc))]>;
1801 def CMPPDrmi : PDIi8<0xC2, MRMSrcMem,
1802 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, SSECC:$cc),
1803 "cmp${cc}pd\t{$src, $dst|$dst, $src}",
1804 [(set VR128:$dst, (int_x86_sse2_cmp_pd VR128:$src1,
1805 (memop addr:$src), imm:$cc))]>;
1807 // Accept explicit immediate argument form instead of comparison code.
1808 let isAsmParserOnly = 1 in {
1809 def CMPPDrri_alt : PDIi8<0xC2, MRMSrcReg,
1810 (outs VR128:$dst), (ins VR128:$src1, VR128:$src, i8imm:$src2),
1811 "cmppd\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
1812 def CMPPDrmi_alt : PDIi8<0xC2, MRMSrcMem,
1813 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src, i8imm:$src2),
1814 "cmppd\t{$src2, $src, $dst|$dst, $src, $src2}", []>;
1817 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), VR128:$src2, imm:$cc)),
1818 (CMPPDrri VR128:$src1, VR128:$src2, imm:$cc)>;
1819 def : Pat<(v2i64 (X86cmppd (v2f64 VR128:$src1), (memop addr:$src2), imm:$cc)),
1820 (CMPPDrmi VR128:$src1, addr:$src2, imm:$cc)>;
1822 // Shuffle and unpack instructions
1823 let Constraints = "$src1 = $dst" in {
1824 def SHUFPDrri : PDIi8<0xC6, MRMSrcReg,
1825 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2, i8imm:$src3),
1826 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1828 (v2f64 (shufp:$src3 VR128:$src1, VR128:$src2)))]>;
1829 def SHUFPDrmi : PDIi8<0xC6, MRMSrcMem,
1830 (outs VR128:$dst), (ins VR128:$src1,
1831 f128mem:$src2, i8imm:$src3),
1832 "shufpd\t{$src3, $src2, $dst|$dst, $src2, $src3}",
1835 VR128:$src1, (memopv2f64 addr:$src2))))]>;
1837 let AddedComplexity = 10 in {
1838 def UNPCKHPDrr : PDI<0x15, MRMSrcReg,
1839 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1840 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1842 (v2f64 (unpckh VR128:$src1, VR128:$src2)))]>;
1843 def UNPCKHPDrm : PDI<0x15, MRMSrcMem,
1844 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1845 "unpckhpd\t{$src2, $dst|$dst, $src2}",
1847 (v2f64 (unpckh VR128:$src1,
1848 (memopv2f64 addr:$src2))))]>;
1850 def UNPCKLPDrr : PDI<0x14, MRMSrcReg,
1851 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
1852 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1854 (v2f64 (unpckl VR128:$src1, VR128:$src2)))]>;
1855 def UNPCKLPDrm : PDI<0x14, MRMSrcMem,
1856 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
1857 "unpcklpd\t{$src2, $dst|$dst, $src2}",
1859 (unpckl VR128:$src1, (memopv2f64 addr:$src2)))]>;
1860 } // AddedComplexity
1861 } // Constraints = "$src1 = $dst"
1864 //===---------------------------------------------------------------------===//
1865 // SSE integer instructions
1866 let ExeDomain = SSEPackedInt in {
1868 // Move Instructions
1869 let neverHasSideEffects = 1 in
1870 def MOVDQArr : PDI<0x6F, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
1871 "movdqa\t{$src, $dst|$dst, $src}", []>;
1872 let canFoldAsLoad = 1, mayLoad = 1 in
1873 def MOVDQArm : PDI<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1874 "movdqa\t{$src, $dst|$dst, $src}",
1875 [/*(set VR128:$dst, (alignedloadv2i64 addr:$src))*/]>;
1877 def MOVDQAmr : PDI<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1878 "movdqa\t{$src, $dst|$dst, $src}",
1879 [/*(alignedstore (v2i64 VR128:$src), addr:$dst)*/]>;
1880 let canFoldAsLoad = 1, mayLoad = 1 in
1881 def MOVDQUrm : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1882 "movdqu\t{$src, $dst|$dst, $src}",
1883 [/*(set VR128:$dst, (loadv2i64 addr:$src))*/]>,
1884 XS, Requires<[HasSSE2]>;
1886 def MOVDQUmr : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1887 "movdqu\t{$src, $dst|$dst, $src}",
1888 [/*(store (v2i64 VR128:$src), addr:$dst)*/]>,
1889 XS, Requires<[HasSSE2]>;
1891 // Intrinsic forms of MOVDQU load and store
1892 let canFoldAsLoad = 1 in
1893 def MOVDQUrm_Int : I<0x6F, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
1894 "movdqu\t{$src, $dst|$dst, $src}",
1895 [(set VR128:$dst, (int_x86_sse2_loadu_dq addr:$src))]>,
1896 XS, Requires<[HasSSE2]>;
1897 def MOVDQUmr_Int : I<0x7F, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
1898 "movdqu\t{$src, $dst|$dst, $src}",
1899 [(int_x86_sse2_storeu_dq addr:$dst, VR128:$src)]>,
1900 XS, Requires<[HasSSE2]>;
1902 let Constraints = "$src1 = $dst" in {
1904 multiclass PDI_binop_rm_int<bits<8> opc, string OpcodeStr, Intrinsic IntId,
1905 bit Commutable = 0> {
1906 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1907 (ins VR128:$src1, VR128:$src2),
1908 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1909 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]> {
1910 let isCommutable = Commutable;
1912 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1913 (ins VR128:$src1, i128mem:$src2),
1914 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1915 [(set VR128:$dst, (IntId VR128:$src1,
1916 (bitconvert (memopv2i64
1920 multiclass PDI_binop_rmi_int<bits<8> opc, bits<8> opc2, Format ImmForm,
1922 Intrinsic IntId, Intrinsic IntId2> {
1923 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1924 (ins VR128:$src1, VR128:$src2),
1925 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1926 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2))]>;
1927 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1928 (ins VR128:$src1, i128mem:$src2),
1929 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1930 [(set VR128:$dst, (IntId VR128:$src1,
1931 (bitconvert (memopv2i64 addr:$src2))))]>;
1932 def ri : PDIi8<opc2, ImmForm, (outs VR128:$dst),
1933 (ins VR128:$src1, i32i8imm:$src2),
1934 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1935 [(set VR128:$dst, (IntId2 VR128:$src1, (i32 imm:$src2)))]>;
1938 /// PDI_binop_rm - Simple SSE2 binary operator.
1939 multiclass PDI_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
1940 ValueType OpVT, bit Commutable = 0> {
1941 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1942 (ins VR128:$src1, VR128:$src2),
1943 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1944 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]> {
1945 let isCommutable = Commutable;
1947 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1948 (ins VR128:$src1, i128mem:$src2),
1949 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1950 [(set VR128:$dst, (OpVT (OpNode VR128:$src1,
1951 (bitconvert (memopv2i64 addr:$src2)))))]>;
1954 /// PDI_binop_rm_v2i64 - Simple SSE2 binary operator whose type is v2i64.
1956 /// FIXME: we could eliminate this and use PDI_binop_rm instead if tblgen knew
1957 /// to collapse (bitconvert VT to VT) into its operand.
1959 multiclass PDI_binop_rm_v2i64<bits<8> opc, string OpcodeStr, SDNode OpNode,
1960 bit Commutable = 0> {
1961 def rr : PDI<opc, MRMSrcReg, (outs VR128:$dst),
1962 (ins VR128:$src1, VR128:$src2),
1963 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1964 [(set VR128:$dst, (v2i64 (OpNode VR128:$src1, VR128:$src2)))]> {
1965 let isCommutable = Commutable;
1967 def rm : PDI<opc, MRMSrcMem, (outs VR128:$dst),
1968 (ins VR128:$src1, i128mem:$src2),
1969 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
1970 [(set VR128:$dst, (OpNode VR128:$src1,
1971 (memopv2i64 addr:$src2)))]>;
1974 } // Constraints = "$src1 = $dst"
1975 } // ExeDomain = SSEPackedInt
1977 // 128-bit Integer Arithmetic
1979 defm PADDB : PDI_binop_rm<0xFC, "paddb", add, v16i8, 1>;
1980 defm PADDW : PDI_binop_rm<0xFD, "paddw", add, v8i16, 1>;
1981 defm PADDD : PDI_binop_rm<0xFE, "paddd", add, v4i32, 1>;
1982 defm PADDQ : PDI_binop_rm_v2i64<0xD4, "paddq", add, 1>;
1984 defm PADDSB : PDI_binop_rm_int<0xEC, "paddsb" , int_x86_sse2_padds_b, 1>;
1985 defm PADDSW : PDI_binop_rm_int<0xED, "paddsw" , int_x86_sse2_padds_w, 1>;
1986 defm PADDUSB : PDI_binop_rm_int<0xDC, "paddusb", int_x86_sse2_paddus_b, 1>;
1987 defm PADDUSW : PDI_binop_rm_int<0xDD, "paddusw", int_x86_sse2_paddus_w, 1>;
1989 defm PSUBB : PDI_binop_rm<0xF8, "psubb", sub, v16i8>;
1990 defm PSUBW : PDI_binop_rm<0xF9, "psubw", sub, v8i16>;
1991 defm PSUBD : PDI_binop_rm<0xFA, "psubd", sub, v4i32>;
1992 defm PSUBQ : PDI_binop_rm_v2i64<0xFB, "psubq", sub>;
1994 defm PSUBSB : PDI_binop_rm_int<0xE8, "psubsb" , int_x86_sse2_psubs_b>;
1995 defm PSUBSW : PDI_binop_rm_int<0xE9, "psubsw" , int_x86_sse2_psubs_w>;
1996 defm PSUBUSB : PDI_binop_rm_int<0xD8, "psubusb", int_x86_sse2_psubus_b>;
1997 defm PSUBUSW : PDI_binop_rm_int<0xD9, "psubusw", int_x86_sse2_psubus_w>;
1999 defm PMULLW : PDI_binop_rm<0xD5, "pmullw", mul, v8i16, 1>;
2001 defm PMULHUW : PDI_binop_rm_int<0xE4, "pmulhuw", int_x86_sse2_pmulhu_w, 1>;
2002 defm PMULHW : PDI_binop_rm_int<0xE5, "pmulhw" , int_x86_sse2_pmulh_w , 1>;
2003 defm PMULUDQ : PDI_binop_rm_int<0xF4, "pmuludq", int_x86_sse2_pmulu_dq, 1>;
2005 defm PMADDWD : PDI_binop_rm_int<0xF5, "pmaddwd", int_x86_sse2_pmadd_wd, 1>;
2007 defm PAVGB : PDI_binop_rm_int<0xE0, "pavgb", int_x86_sse2_pavg_b, 1>;
2008 defm PAVGW : PDI_binop_rm_int<0xE3, "pavgw", int_x86_sse2_pavg_w, 1>;
2011 defm PMINUB : PDI_binop_rm_int<0xDA, "pminub", int_x86_sse2_pminu_b, 1>;
2012 defm PMINSW : PDI_binop_rm_int<0xEA, "pminsw", int_x86_sse2_pmins_w, 1>;
2013 defm PMAXUB : PDI_binop_rm_int<0xDE, "pmaxub", int_x86_sse2_pmaxu_b, 1>;
2014 defm PMAXSW : PDI_binop_rm_int<0xEE, "pmaxsw", int_x86_sse2_pmaxs_w, 1>;
2015 defm PSADBW : PDI_binop_rm_int<0xF6, "psadbw", int_x86_sse2_psad_bw, 1>;
2018 defm PSLLW : PDI_binop_rmi_int<0xF1, 0x71, MRM6r, "psllw",
2019 int_x86_sse2_psll_w, int_x86_sse2_pslli_w>;
2020 defm PSLLD : PDI_binop_rmi_int<0xF2, 0x72, MRM6r, "pslld",
2021 int_x86_sse2_psll_d, int_x86_sse2_pslli_d>;
2022 defm PSLLQ : PDI_binop_rmi_int<0xF3, 0x73, MRM6r, "psllq",
2023 int_x86_sse2_psll_q, int_x86_sse2_pslli_q>;
2025 defm PSRLW : PDI_binop_rmi_int<0xD1, 0x71, MRM2r, "psrlw",
2026 int_x86_sse2_psrl_w, int_x86_sse2_psrli_w>;
2027 defm PSRLD : PDI_binop_rmi_int<0xD2, 0x72, MRM2r, "psrld",
2028 int_x86_sse2_psrl_d, int_x86_sse2_psrli_d>;
2029 defm PSRLQ : PDI_binop_rmi_int<0xD3, 0x73, MRM2r, "psrlq",
2030 int_x86_sse2_psrl_q, int_x86_sse2_psrli_q>;
2032 defm PSRAW : PDI_binop_rmi_int<0xE1, 0x71, MRM4r, "psraw",
2033 int_x86_sse2_psra_w, int_x86_sse2_psrai_w>;
2034 defm PSRAD : PDI_binop_rmi_int<0xE2, 0x72, MRM4r, "psrad",
2035 int_x86_sse2_psra_d, int_x86_sse2_psrai_d>;
2037 // 128-bit logical shifts.
2038 let Constraints = "$src1 = $dst", neverHasSideEffects = 1,
2039 ExeDomain = SSEPackedInt in {
2040 def PSLLDQri : PDIi8<0x73, MRM7r,
2041 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
2042 "pslldq\t{$src2, $dst|$dst, $src2}", []>;
2043 def PSRLDQri : PDIi8<0x73, MRM3r,
2044 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
2045 "psrldq\t{$src2, $dst|$dst, $src2}", []>;
2046 // PSRADQri doesn't exist in SSE[1-3].
2049 let Predicates = [HasSSE2] in {
2050 def : Pat<(int_x86_sse2_psll_dq VR128:$src1, imm:$src2),
2051 (v2i64 (PSLLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2052 def : Pat<(int_x86_sse2_psrl_dq VR128:$src1, imm:$src2),
2053 (v2i64 (PSRLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2054 def : Pat<(int_x86_sse2_psll_dq_bs VR128:$src1, imm:$src2),
2055 (v2i64 (PSLLDQri VR128:$src1, imm:$src2))>;
2056 def : Pat<(int_x86_sse2_psrl_dq_bs VR128:$src1, imm:$src2),
2057 (v2i64 (PSRLDQri VR128:$src1, imm:$src2))>;
2058 def : Pat<(v2f64 (X86fsrl VR128:$src1, i32immSExt8:$src2)),
2059 (v2f64 (PSRLDQri VR128:$src1, (BYTE_imm imm:$src2)))>;
2061 // Shift up / down and insert zero's.
2062 def : Pat<(v2i64 (X86vshl VR128:$src, (i8 imm:$amt))),
2063 (v2i64 (PSLLDQri VR128:$src, (BYTE_imm imm:$amt)))>;
2064 def : Pat<(v2i64 (X86vshr VR128:$src, (i8 imm:$amt))),
2065 (v2i64 (PSRLDQri VR128:$src, (BYTE_imm imm:$amt)))>;
2069 defm PAND : PDI_binop_rm_v2i64<0xDB, "pand", and, 1>;
2070 defm POR : PDI_binop_rm_v2i64<0xEB, "por" , or , 1>;
2071 defm PXOR : PDI_binop_rm_v2i64<0xEF, "pxor", xor, 1>;
2073 let Constraints = "$src1 = $dst", ExeDomain = SSEPackedInt in {
2074 def PANDNrr : PDI<0xDF, MRMSrcReg,
2075 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2076 "pandn\t{$src2, $dst|$dst, $src2}",
2077 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
2080 def PANDNrm : PDI<0xDF, MRMSrcMem,
2081 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2082 "pandn\t{$src2, $dst|$dst, $src2}",
2083 [(set VR128:$dst, (v2i64 (and (vnot VR128:$src1),
2084 (memopv2i64 addr:$src2))))]>;
2087 // SSE2 Integer comparison
2088 defm PCMPEQB : PDI_binop_rm_int<0x74, "pcmpeqb", int_x86_sse2_pcmpeq_b>;
2089 defm PCMPEQW : PDI_binop_rm_int<0x75, "pcmpeqw", int_x86_sse2_pcmpeq_w>;
2090 defm PCMPEQD : PDI_binop_rm_int<0x76, "pcmpeqd", int_x86_sse2_pcmpeq_d>;
2091 defm PCMPGTB : PDI_binop_rm_int<0x64, "pcmpgtb", int_x86_sse2_pcmpgt_b>;
2092 defm PCMPGTW : PDI_binop_rm_int<0x65, "pcmpgtw", int_x86_sse2_pcmpgt_w>;
2093 defm PCMPGTD : PDI_binop_rm_int<0x66, "pcmpgtd", int_x86_sse2_pcmpgt_d>;
2095 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, VR128:$src2)),
2096 (PCMPEQBrr VR128:$src1, VR128:$src2)>;
2097 def : Pat<(v16i8 (X86pcmpeqb VR128:$src1, (memop addr:$src2))),
2098 (PCMPEQBrm VR128:$src1, addr:$src2)>;
2099 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, VR128:$src2)),
2100 (PCMPEQWrr VR128:$src1, VR128:$src2)>;
2101 def : Pat<(v8i16 (X86pcmpeqw VR128:$src1, (memop addr:$src2))),
2102 (PCMPEQWrm VR128:$src1, addr:$src2)>;
2103 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, VR128:$src2)),
2104 (PCMPEQDrr VR128:$src1, VR128:$src2)>;
2105 def : Pat<(v4i32 (X86pcmpeqd VR128:$src1, (memop addr:$src2))),
2106 (PCMPEQDrm VR128:$src1, addr:$src2)>;
2108 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, VR128:$src2)),
2109 (PCMPGTBrr VR128:$src1, VR128:$src2)>;
2110 def : Pat<(v16i8 (X86pcmpgtb VR128:$src1, (memop addr:$src2))),
2111 (PCMPGTBrm VR128:$src1, addr:$src2)>;
2112 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, VR128:$src2)),
2113 (PCMPGTWrr VR128:$src1, VR128:$src2)>;
2114 def : Pat<(v8i16 (X86pcmpgtw VR128:$src1, (memop addr:$src2))),
2115 (PCMPGTWrm VR128:$src1, addr:$src2)>;
2116 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, VR128:$src2)),
2117 (PCMPGTDrr VR128:$src1, VR128:$src2)>;
2118 def : Pat<(v4i32 (X86pcmpgtd VR128:$src1, (memop addr:$src2))),
2119 (PCMPGTDrm VR128:$src1, addr:$src2)>;
2122 // Pack instructions
2123 defm PACKSSWB : PDI_binop_rm_int<0x63, "packsswb", int_x86_sse2_packsswb_128>;
2124 defm PACKSSDW : PDI_binop_rm_int<0x6B, "packssdw", int_x86_sse2_packssdw_128>;
2125 defm PACKUSWB : PDI_binop_rm_int<0x67, "packuswb", int_x86_sse2_packuswb_128>;
2127 let ExeDomain = SSEPackedInt in {
2129 // Shuffle and unpack instructions
2130 let AddedComplexity = 5 in {
2131 def PSHUFDri : PDIi8<0x70, MRMSrcReg,
2132 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2133 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2134 [(set VR128:$dst, (v4i32 (pshufd:$src2
2135 VR128:$src1, (undef))))]>;
2136 def PSHUFDmi : PDIi8<0x70, MRMSrcMem,
2137 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2138 "pshufd\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2139 [(set VR128:$dst, (v4i32 (pshufd:$src2
2140 (bc_v4i32 (memopv2i64 addr:$src1)),
2144 // SSE2 with ImmT == Imm8 and XS prefix.
2145 def PSHUFHWri : Ii8<0x70, MRMSrcReg,
2146 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2147 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2148 [(set VR128:$dst, (v8i16 (pshufhw:$src2 VR128:$src1,
2150 XS, Requires<[HasSSE2]>;
2151 def PSHUFHWmi : Ii8<0x70, MRMSrcMem,
2152 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2153 "pshufhw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2154 [(set VR128:$dst, (v8i16 (pshufhw:$src2
2155 (bc_v8i16 (memopv2i64 addr:$src1)),
2157 XS, Requires<[HasSSE2]>;
2159 // SSE2 with ImmT == Imm8 and XD prefix.
2160 def PSHUFLWri : Ii8<0x70, MRMSrcReg,
2161 (outs VR128:$dst), (ins VR128:$src1, i8imm:$src2),
2162 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2163 [(set VR128:$dst, (v8i16 (pshuflw:$src2 VR128:$src1,
2165 XD, Requires<[HasSSE2]>;
2166 def PSHUFLWmi : Ii8<0x70, MRMSrcMem,
2167 (outs VR128:$dst), (ins i128mem:$src1, i8imm:$src2),
2168 "pshuflw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2169 [(set VR128:$dst, (v8i16 (pshuflw:$src2
2170 (bc_v8i16 (memopv2i64 addr:$src1)),
2172 XD, Requires<[HasSSE2]>;
2174 // Unpack instructions
2175 multiclass sse2_unpack<bits<8> opc, string OpcodeStr, ValueType vt,
2176 PatFrag unp_frag, PatFrag bc_frag> {
2177 def rr : PDI<opc, MRMSrcReg,
2178 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2179 !strconcat(OpcodeStr,"\t{$src2, $dst|$dst, $src2}"),
2180 [(set VR128:$dst, (vt (unp_frag VR128:$src1, VR128:$src2)))]>;
2181 def rm : PDI<opc, MRMSrcMem,
2182 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2183 !strconcat(OpcodeStr,"\t{$src2, $dst|$dst, $src2}"),
2184 [(set VR128:$dst, (unp_frag VR128:$src1,
2185 (bc_frag (memopv2i64
2189 let Constraints = "$src1 = $dst" in {
2190 defm PUNPCKLBW : sse2_unpack<0x60, "punpcklbw", v16i8, unpckl, bc_v16i8>;
2191 defm PUNPCKLWD : sse2_unpack<0x61, "punpcklwd", v8i16, unpckl, bc_v8i16>;
2192 defm PUNPCKLDQ : sse2_unpack<0x62, "punpckldq", v4i32, unpckl, bc_v4i32>;
2194 /// FIXME: we could eliminate this and use sse2_unpack instead if tblgen
2195 /// knew to collapse (bitconvert VT to VT) into its operand.
2196 def PUNPCKLQDQrr : PDI<0x6C, MRMSrcReg,
2197 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2198 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2200 (v2i64 (unpckl VR128:$src1, VR128:$src2)))]>;
2201 def PUNPCKLQDQrm : PDI<0x6C, MRMSrcMem,
2202 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2203 "punpcklqdq\t{$src2, $dst|$dst, $src2}",
2205 (v2i64 (unpckl VR128:$src1,
2206 (memopv2i64 addr:$src2))))]>;
2208 defm PUNPCKHBW : sse2_unpack<0x68, "punpckhbw", v16i8, unpckh, bc_v16i8>;
2209 defm PUNPCKHWD : sse2_unpack<0x69, "punpckhwd", v8i16, unpckh, bc_v8i16>;
2210 defm PUNPCKHDQ : sse2_unpack<0x6A, "punpckhdq", v4i32, unpckh, bc_v4i32>;
2212 /// FIXME: we could eliminate this and use sse2_unpack instead if tblgen
2213 /// knew to collapse (bitconvert VT to VT) into its operand.
2214 def PUNPCKHQDQrr : PDI<0x6D, MRMSrcReg,
2215 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2216 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2218 (v2i64 (unpckh VR128:$src1, VR128:$src2)))]>;
2219 def PUNPCKHQDQrm : PDI<0x6D, MRMSrcMem,
2220 (outs VR128:$dst), (ins VR128:$src1, i128mem:$src2),
2221 "punpckhqdq\t{$src2, $dst|$dst, $src2}",
2223 (v2i64 (unpckh VR128:$src1,
2224 (memopv2i64 addr:$src2))))]>;
2228 def PEXTRWri : PDIi8<0xC5, MRMSrcReg,
2229 (outs GR32:$dst), (ins VR128:$src1, i32i8imm:$src2),
2230 "pextrw\t{$src2, $src1, $dst|$dst, $src1, $src2}",
2231 [(set GR32:$dst, (X86pextrw (v8i16 VR128:$src1),
2233 let Constraints = "$src1 = $dst" in {
2234 def PINSRWrri : PDIi8<0xC4, MRMSrcReg,
2235 (outs VR128:$dst), (ins VR128:$src1,
2236 GR32:$src2, i32i8imm:$src3),
2237 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2239 (X86pinsrw VR128:$src1, GR32:$src2, imm:$src3))]>;
2240 def PINSRWrmi : PDIi8<0xC4, MRMSrcMem,
2241 (outs VR128:$dst), (ins VR128:$src1,
2242 i16mem:$src2, i32i8imm:$src3),
2243 "pinsrw\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2245 (X86pinsrw VR128:$src1, (extloadi16 addr:$src2),
2250 def PMOVMSKBrr : PDI<0xD7, MRMSrcReg, (outs GR32:$dst), (ins VR128:$src),
2251 "pmovmskb\t{$src, $dst|$dst, $src}",
2252 [(set GR32:$dst, (int_x86_sse2_pmovmskb_128 VR128:$src))]>;
2254 // Conditional store
2256 def MASKMOVDQU : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2257 "maskmovdqu\t{$mask, $src|$src, $mask}",
2258 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, EDI)]>;
2261 def MASKMOVDQU64 : PDI<0xF7, MRMSrcReg, (outs), (ins VR128:$src, VR128:$mask),
2262 "maskmovdqu\t{$mask, $src|$src, $mask}",
2263 [(int_x86_sse2_maskmov_dqu VR128:$src, VR128:$mask, RDI)]>;
2265 } // ExeDomain = SSEPackedInt
2267 // Non-temporal stores
2268 def MOVNTPDmr_Int : PDI<0x2B, MRMDestMem, (outs), (ins i128mem:$dst, VR128:$src),
2269 "movntpd\t{$src, $dst|$dst, $src}",
2270 [(int_x86_sse2_movnt_pd addr:$dst, VR128:$src)]>;
2271 let ExeDomain = SSEPackedInt in
2272 def MOVNTDQmr_Int : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2273 "movntdq\t{$src, $dst|$dst, $src}",
2274 [(int_x86_sse2_movnt_dq addr:$dst, VR128:$src)]>;
2275 def MOVNTImr_Int : I<0xC3, MRMDestMem, (outs), (ins i32mem:$dst, GR32:$src),
2276 "movnti\t{$src, $dst|$dst, $src}",
2277 [(int_x86_sse2_movnt_i addr:$dst, GR32:$src)]>,
2278 TB, Requires<[HasSSE2]>;
2280 let AddedComplexity = 400 in { // Prefer non-temporal versions
2281 def MOVNTPDmr : PDI<0x2B, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2282 "movntpd\t{$src, $dst|$dst, $src}",
2283 [(alignednontemporalstore(v2f64 VR128:$src), addr:$dst)]>;
2285 let ExeDomain = SSEPackedInt in
2286 def MOVNTDQmr : PDI<0xE7, MRMDestMem, (outs), (ins f128mem:$dst, VR128:$src),
2287 "movntdq\t{$src, $dst|$dst, $src}",
2288 [(alignednontemporalstore (v4f32 VR128:$src), addr:$dst)]>;
2292 def CLFLUSH : I<0xAE, MRM7m, (outs), (ins i8mem:$src),
2293 "clflush\t$src", [(int_x86_sse2_clflush addr:$src)]>,
2294 TB, Requires<[HasSSE2]>;
2296 // Load, store, and memory fence
2297 def LFENCE : I<0xAE, MRM_E8, (outs), (ins),
2298 "lfence", [(int_x86_sse2_lfence)]>, TB, Requires<[HasSSE2]>;
2299 def MFENCE : I<0xAE, MRM_F0, (outs), (ins),
2300 "mfence", [(int_x86_sse2_mfence)]>, TB, Requires<[HasSSE2]>;
2302 // Pause. This "instruction" is encoded as "rep; nop", so even though it
2303 // was introduced with SSE2, it's backward compatible.
2304 def PAUSE : I<0x90, RawFrm, (outs), (ins), "pause", []>, REP;
2306 //TODO: custom lower this so as to never even generate the noop
2307 def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm),
2309 def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
2310 def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
2311 def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm),
2314 // Alias instructions that map zero vector to pxor / xorp* for sse.
2315 // We set canFoldAsLoad because this can be converted to a constant-pool
2316 // load of an all-ones value if folding it would be beneficial.
2317 let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
2318 isCodeGenOnly = 1, ExeDomain = SSEPackedInt in
2319 // FIXME: Change encoding to pseudo.
2320 def V_SETALLONES : PDI<0x76, MRMInitReg, (outs VR128:$dst), (ins), "",
2321 [(set VR128:$dst, (v4i32 immAllOnesV))]>;
2323 def MOVDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2324 "movd\t{$src, $dst|$dst, $src}",
2326 (v4i32 (scalar_to_vector GR32:$src)))]>;
2327 def MOVDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2328 "movd\t{$src, $dst|$dst, $src}",
2330 (v4i32 (scalar_to_vector (loadi32 addr:$src))))]>;
2332 def MOVDI2SSrr : PDI<0x6E, MRMSrcReg, (outs FR32:$dst), (ins GR32:$src),
2333 "movd\t{$src, $dst|$dst, $src}",
2334 [(set FR32:$dst, (bitconvert GR32:$src))]>;
2336 def MOVDI2SSrm : PDI<0x6E, MRMSrcMem, (outs FR32:$dst), (ins i32mem:$src),
2337 "movd\t{$src, $dst|$dst, $src}",
2338 [(set FR32:$dst, (bitconvert (loadi32 addr:$src)))]>;
2340 // SSE2 instructions with XS prefix
2341 def MOVQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2342 "movq\t{$src, $dst|$dst, $src}",
2344 (v2i64 (scalar_to_vector (loadi64 addr:$src))))]>, XS,
2345 Requires<[HasSSE2]>;
2346 def MOVPQI2QImr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2347 "movq\t{$src, $dst|$dst, $src}",
2348 [(store (i64 (vector_extract (v2i64 VR128:$src),
2349 (iPTR 0))), addr:$dst)]>;
2351 def : Pat<(f64 (vector_extract (v2f64 VR128:$src), (iPTR 0))),
2352 (f64 (EXTRACT_SUBREG (v2f64 VR128:$src), sub_sd))>;
2354 def MOVPDI2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins VR128:$src),
2355 "movd\t{$src, $dst|$dst, $src}",
2356 [(set GR32:$dst, (vector_extract (v4i32 VR128:$src),
2358 def MOVPDI2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, VR128:$src),
2359 "movd\t{$src, $dst|$dst, $src}",
2360 [(store (i32 (vector_extract (v4i32 VR128:$src),
2361 (iPTR 0))), addr:$dst)]>;
2363 def MOVSS2DIrr : PDI<0x7E, MRMDestReg, (outs GR32:$dst), (ins FR32:$src),
2364 "movd\t{$src, $dst|$dst, $src}",
2365 [(set GR32:$dst, (bitconvert FR32:$src))]>;
2366 def MOVSS2DImr : PDI<0x7E, MRMDestMem, (outs), (ins i32mem:$dst, FR32:$src),
2367 "movd\t{$src, $dst|$dst, $src}",
2368 [(store (i32 (bitconvert FR32:$src)), addr:$dst)]>;
2370 // Store / copy lower 64-bits of a XMM register.
2371 def MOVLQ128mr : PDI<0xD6, MRMDestMem, (outs), (ins i64mem:$dst, VR128:$src),
2372 "movq\t{$src, $dst|$dst, $src}",
2373 [(int_x86_sse2_storel_dq addr:$dst, VR128:$src)]>;
2375 // movd / movq to XMM register zero-extends
2376 let AddedComplexity = 15 in {
2377 def MOVZDI2PDIrr : PDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR32:$src),
2378 "movd\t{$src, $dst|$dst, $src}",
2379 [(set VR128:$dst, (v4i32 (X86vzmovl
2380 (v4i32 (scalar_to_vector GR32:$src)))))]>;
2381 // This is X86-64 only.
2382 def MOVZQI2PQIrr : RPDI<0x6E, MRMSrcReg, (outs VR128:$dst), (ins GR64:$src),
2383 "mov{d|q}\t{$src, $dst|$dst, $src}",
2384 [(set VR128:$dst, (v2i64 (X86vzmovl
2385 (v2i64 (scalar_to_vector GR64:$src)))))]>;
2388 let AddedComplexity = 20 in {
2389 def MOVZDI2PDIrm : PDI<0x6E, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
2390 "movd\t{$src, $dst|$dst, $src}",
2392 (v4i32 (X86vzmovl (v4i32 (scalar_to_vector
2393 (loadi32 addr:$src))))))]>;
2395 def : Pat<(v4i32 (X86vzmovl (loadv4i32 addr:$src))),
2396 (MOVZDI2PDIrm addr:$src)>;
2397 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv4f32 addr:$src)))),
2398 (MOVZDI2PDIrm addr:$src)>;
2399 def : Pat<(v4i32 (X86vzmovl (bc_v4i32 (loadv2i64 addr:$src)))),
2400 (MOVZDI2PDIrm addr:$src)>;
2402 def MOVZQI2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
2403 "movq\t{$src, $dst|$dst, $src}",
2405 (v2i64 (X86vzmovl (v2i64 (scalar_to_vector
2406 (loadi64 addr:$src))))))]>, XS,
2407 Requires<[HasSSE2]>;
2409 def : Pat<(v2i64 (X86vzmovl (loadv2i64 addr:$src))),
2410 (MOVZQI2PQIrm addr:$src)>;
2411 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4f32 addr:$src)))),
2412 (MOVZQI2PQIrm addr:$src)>;
2413 def : Pat<(v2i64 (X86vzload addr:$src)), (MOVZQI2PQIrm addr:$src)>;
2416 // Moving from XMM to XMM and clear upper 64 bits. Note, there is a bug in
2417 // IA32 document. movq xmm1, xmm2 does clear the high bits.
2418 let AddedComplexity = 15 in
2419 def MOVZPQILo2PQIrr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2420 "movq\t{$src, $dst|$dst, $src}",
2421 [(set VR128:$dst, (v2i64 (X86vzmovl (v2i64 VR128:$src))))]>,
2422 XS, Requires<[HasSSE2]>;
2424 let AddedComplexity = 20 in {
2425 def MOVZPQILo2PQIrm : I<0x7E, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2426 "movq\t{$src, $dst|$dst, $src}",
2427 [(set VR128:$dst, (v2i64 (X86vzmovl
2428 (loadv2i64 addr:$src))))]>,
2429 XS, Requires<[HasSSE2]>;
2431 def : Pat<(v2i64 (X86vzmovl (bc_v2i64 (loadv4i32 addr:$src)))),
2432 (MOVZPQILo2PQIrm addr:$src)>;
2435 // Instructions for the disassembler
2436 // xr = XMM register
2439 def MOVQxrxr : I<0x7E, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2440 "movq\t{$src, $dst|$dst, $src}", []>, XS;
2442 //===---------------------------------------------------------------------===//
2443 // SSE3 Instructions
2444 //===---------------------------------------------------------------------===//
2446 // Move Instructions
2447 def MOVSHDUPrr : S3SI<0x16, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2448 "movshdup\t{$src, $dst|$dst, $src}",
2449 [(set VR128:$dst, (v4f32 (movshdup
2450 VR128:$src, (undef))))]>;
2451 def MOVSHDUPrm : S3SI<0x16, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2452 "movshdup\t{$src, $dst|$dst, $src}",
2453 [(set VR128:$dst, (movshdup
2454 (memopv4f32 addr:$src), (undef)))]>;
2456 def MOVSLDUPrr : S3SI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2457 "movsldup\t{$src, $dst|$dst, $src}",
2458 [(set VR128:$dst, (v4f32 (movsldup
2459 VR128:$src, (undef))))]>;
2460 def MOVSLDUPrm : S3SI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f128mem:$src),
2461 "movsldup\t{$src, $dst|$dst, $src}",
2462 [(set VR128:$dst, (movsldup
2463 (memopv4f32 addr:$src), (undef)))]>;
2465 def MOVDDUPrr : S3DI<0x12, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
2466 "movddup\t{$src, $dst|$dst, $src}",
2467 [(set VR128:$dst,(v2f64 (movddup VR128:$src, (undef))))]>;
2468 def MOVDDUPrm : S3DI<0x12, MRMSrcMem, (outs VR128:$dst), (ins f64mem:$src),
2469 "movddup\t{$src, $dst|$dst, $src}",
2471 (v2f64 (movddup (scalar_to_vector (loadf64 addr:$src)),
2474 def : Pat<(movddup (bc_v2f64 (v2i64 (scalar_to_vector (loadi64 addr:$src)))),
2476 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2478 let AddedComplexity = 5 in {
2479 def : Pat<(movddup (memopv2f64 addr:$src), (undef)),
2480 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2481 def : Pat<(movddup (bc_v4f32 (memopv2f64 addr:$src)), (undef)),
2482 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2483 def : Pat<(movddup (memopv2i64 addr:$src), (undef)),
2484 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2485 def : Pat<(movddup (bc_v4i32 (memopv2i64 addr:$src)), (undef)),
2486 (MOVDDUPrm addr:$src)>, Requires<[HasSSE3]>;
2490 let Constraints = "$src1 = $dst" in {
2491 def ADDSUBPSrr : S3DI<0xD0, MRMSrcReg,
2492 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2493 "addsubps\t{$src2, $dst|$dst, $src2}",
2494 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2496 def ADDSUBPSrm : S3DI<0xD0, MRMSrcMem,
2497 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2498 "addsubps\t{$src2, $dst|$dst, $src2}",
2499 [(set VR128:$dst, (int_x86_sse3_addsub_ps VR128:$src1,
2500 (memop addr:$src2)))]>;
2501 def ADDSUBPDrr : S3I<0xD0, MRMSrcReg,
2502 (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2503 "addsubpd\t{$src2, $dst|$dst, $src2}",
2504 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2506 def ADDSUBPDrm : S3I<0xD0, MRMSrcMem,
2507 (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2508 "addsubpd\t{$src2, $dst|$dst, $src2}",
2509 [(set VR128:$dst, (int_x86_sse3_addsub_pd VR128:$src1,
2510 (memop addr:$src2)))]>;
2513 def LDDQUrm : S3DI<0xF0, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
2514 "lddqu\t{$src, $dst|$dst, $src}",
2515 [(set VR128:$dst, (int_x86_sse3_ldu_dq addr:$src))]>;
2518 class S3D_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2519 : S3DI<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2520 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2521 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, VR128:$src2)))]>;
2522 class S3D_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2523 : S3DI<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2524 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2525 [(set VR128:$dst, (v4f32 (IntId VR128:$src1, (memop addr:$src2))))]>;
2526 class S3_Intrr<bits<8> o, string OpcodeStr, Intrinsic IntId>
2527 : S3I<o, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src1, VR128:$src2),
2528 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2529 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, VR128:$src2)))]>;
2530 class S3_Intrm<bits<8> o, string OpcodeStr, Intrinsic IntId>
2531 : S3I<o, MRMSrcMem, (outs VR128:$dst), (ins VR128:$src1, f128mem:$src2),
2532 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2533 [(set VR128:$dst, (v2f64 (IntId VR128:$src1, (memopv2f64 addr:$src2))))]>;
2535 let Constraints = "$src1 = $dst" in {
2536 def HADDPSrr : S3D_Intrr<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2537 def HADDPSrm : S3D_Intrm<0x7C, "haddps", int_x86_sse3_hadd_ps>;
2538 def HADDPDrr : S3_Intrr <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2539 def HADDPDrm : S3_Intrm <0x7C, "haddpd", int_x86_sse3_hadd_pd>;
2540 def HSUBPSrr : S3D_Intrr<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2541 def HSUBPSrm : S3D_Intrm<0x7D, "hsubps", int_x86_sse3_hsub_ps>;
2542 def HSUBPDrr : S3_Intrr <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2543 def HSUBPDrm : S3_Intrm <0x7D, "hsubpd", int_x86_sse3_hsub_pd>;
2546 // Thread synchronization
2547 def MONITOR : I<0x01, MRM_C8, (outs), (ins), "monitor",
2548 [(int_x86_sse3_monitor EAX, ECX, EDX)]>,TB, Requires<[HasSSE3]>;
2549 def MWAIT : I<0x01, MRM_C9, (outs), (ins), "mwait",
2550 [(int_x86_sse3_mwait ECX, EAX)]>, TB, Requires<[HasSSE3]>;
2552 // vector_shuffle v1, <undef> <1, 1, 3, 3>
2553 let AddedComplexity = 15 in
2554 def : Pat<(v4i32 (movshdup VR128:$src, (undef))),
2555 (MOVSHDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2556 let AddedComplexity = 20 in
2557 def : Pat<(v4i32 (movshdup (bc_v4i32 (memopv2i64 addr:$src)), (undef))),
2558 (MOVSHDUPrm addr:$src)>, Requires<[HasSSE3]>;
2560 // vector_shuffle v1, <undef> <0, 0, 2, 2>
2561 let AddedComplexity = 15 in
2562 def : Pat<(v4i32 (movsldup VR128:$src, (undef))),
2563 (MOVSLDUPrr VR128:$src)>, Requires<[HasSSE3]>;
2564 let AddedComplexity = 20 in
2565 def : Pat<(v4i32 (movsldup (bc_v4i32 (memopv2i64 addr:$src)), (undef))),
2566 (MOVSLDUPrm addr:$src)>, Requires<[HasSSE3]>;
2568 //===---------------------------------------------------------------------===//
2569 // SSSE3 Instructions
2570 //===---------------------------------------------------------------------===//
2572 /// SS3I_unop_rm_int_8 - Simple SSSE3 unary operator whose type is v*i8.
2573 multiclass SS3I_unop_rm_int_8<bits<8> opc, string OpcodeStr,
2574 Intrinsic IntId64, Intrinsic IntId128> {
2575 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst), (ins VR64:$src),
2576 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2577 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2579 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst), (ins i64mem:$src),
2580 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2582 (IntId64 (bitconvert (memopv8i8 addr:$src))))]>;
2584 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2586 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2587 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2590 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2592 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2595 (bitconvert (memopv16i8 addr:$src))))]>, OpSize;
2598 /// SS3I_unop_rm_int_16 - Simple SSSE3 unary operator whose type is v*i16.
2599 multiclass SS3I_unop_rm_int_16<bits<8> opc, string OpcodeStr,
2600 Intrinsic IntId64, Intrinsic IntId128> {
2601 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2603 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2604 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2606 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2608 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2611 (bitconvert (memopv4i16 addr:$src))))]>;
2613 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2615 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2616 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2619 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2621 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2624 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
2627 /// SS3I_unop_rm_int_32 - Simple SSSE3 unary operator whose type is v*i32.
2628 multiclass SS3I_unop_rm_int_32<bits<8> opc, string OpcodeStr,
2629 Intrinsic IntId64, Intrinsic IntId128> {
2630 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2632 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2633 [(set VR64:$dst, (IntId64 VR64:$src))]>;
2635 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2637 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2640 (bitconvert (memopv2i32 addr:$src))))]>;
2642 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2644 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2645 [(set VR128:$dst, (IntId128 VR128:$src))]>,
2648 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2650 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
2653 (bitconvert (memopv4i32 addr:$src))))]>, OpSize;
2656 defm PABSB : SS3I_unop_rm_int_8 <0x1C, "pabsb",
2657 int_x86_ssse3_pabs_b,
2658 int_x86_ssse3_pabs_b_128>;
2659 defm PABSW : SS3I_unop_rm_int_16<0x1D, "pabsw",
2660 int_x86_ssse3_pabs_w,
2661 int_x86_ssse3_pabs_w_128>;
2662 defm PABSD : SS3I_unop_rm_int_32<0x1E, "pabsd",
2663 int_x86_ssse3_pabs_d,
2664 int_x86_ssse3_pabs_d_128>;
2666 /// SS3I_binop_rm_int_8 - Simple SSSE3 binary operator whose type is v*i8.
2667 let Constraints = "$src1 = $dst" in {
2668 multiclass SS3I_binop_rm_int_8<bits<8> opc, string OpcodeStr,
2669 Intrinsic IntId64, Intrinsic IntId128,
2670 bit Commutable = 0> {
2671 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2672 (ins VR64:$src1, VR64:$src2),
2673 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2674 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2675 let isCommutable = Commutable;
2677 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2678 (ins VR64:$src1, i64mem:$src2),
2679 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2681 (IntId64 VR64:$src1,
2682 (bitconvert (memopv8i8 addr:$src2))))]>;
2684 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2685 (ins VR128:$src1, VR128:$src2),
2686 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2687 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2689 let isCommutable = Commutable;
2691 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2692 (ins VR128:$src1, i128mem:$src2),
2693 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2695 (IntId128 VR128:$src1,
2696 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
2700 /// SS3I_binop_rm_int_16 - Simple SSSE3 binary operator whose type is v*i16.
2701 let Constraints = "$src1 = $dst" in {
2702 multiclass SS3I_binop_rm_int_16<bits<8> opc, string OpcodeStr,
2703 Intrinsic IntId64, Intrinsic IntId128,
2704 bit Commutable = 0> {
2705 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2706 (ins VR64:$src1, VR64:$src2),
2707 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2708 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2709 let isCommutable = Commutable;
2711 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2712 (ins VR64:$src1, i64mem:$src2),
2713 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2715 (IntId64 VR64:$src1,
2716 (bitconvert (memopv4i16 addr:$src2))))]>;
2718 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2719 (ins VR128:$src1, VR128:$src2),
2720 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2721 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2723 let isCommutable = Commutable;
2725 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2726 (ins VR128:$src1, i128mem:$src2),
2727 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2729 (IntId128 VR128:$src1,
2730 (bitconvert (memopv8i16 addr:$src2))))]>, OpSize;
2734 /// SS3I_binop_rm_int_32 - Simple SSSE3 binary operator whose type is v*i32.
2735 let Constraints = "$src1 = $dst" in {
2736 multiclass SS3I_binop_rm_int_32<bits<8> opc, string OpcodeStr,
2737 Intrinsic IntId64, Intrinsic IntId128,
2738 bit Commutable = 0> {
2739 def rr64 : SS38I<opc, MRMSrcReg, (outs VR64:$dst),
2740 (ins VR64:$src1, VR64:$src2),
2741 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2742 [(set VR64:$dst, (IntId64 VR64:$src1, VR64:$src2))]> {
2743 let isCommutable = Commutable;
2745 def rm64 : SS38I<opc, MRMSrcMem, (outs VR64:$dst),
2746 (ins VR64:$src1, i64mem:$src2),
2747 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2749 (IntId64 VR64:$src1,
2750 (bitconvert (memopv2i32 addr:$src2))))]>;
2752 def rr128 : SS38I<opc, MRMSrcReg, (outs VR128:$dst),
2753 (ins VR128:$src1, VR128:$src2),
2754 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2755 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
2757 let isCommutable = Commutable;
2759 def rm128 : SS38I<opc, MRMSrcMem, (outs VR128:$dst),
2760 (ins VR128:$src1, i128mem:$src2),
2761 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
2763 (IntId128 VR128:$src1,
2764 (bitconvert (memopv4i32 addr:$src2))))]>, OpSize;
2768 let ImmT = NoImm in { // None of these have i8 immediate fields.
2769 defm PHADDW : SS3I_binop_rm_int_16<0x01, "phaddw",
2770 int_x86_ssse3_phadd_w,
2771 int_x86_ssse3_phadd_w_128>;
2772 defm PHADDD : SS3I_binop_rm_int_32<0x02, "phaddd",
2773 int_x86_ssse3_phadd_d,
2774 int_x86_ssse3_phadd_d_128>;
2775 defm PHADDSW : SS3I_binop_rm_int_16<0x03, "phaddsw",
2776 int_x86_ssse3_phadd_sw,
2777 int_x86_ssse3_phadd_sw_128>;
2778 defm PHSUBW : SS3I_binop_rm_int_16<0x05, "phsubw",
2779 int_x86_ssse3_phsub_w,
2780 int_x86_ssse3_phsub_w_128>;
2781 defm PHSUBD : SS3I_binop_rm_int_32<0x06, "phsubd",
2782 int_x86_ssse3_phsub_d,
2783 int_x86_ssse3_phsub_d_128>;
2784 defm PHSUBSW : SS3I_binop_rm_int_16<0x07, "phsubsw",
2785 int_x86_ssse3_phsub_sw,
2786 int_x86_ssse3_phsub_sw_128>;
2787 defm PMADDUBSW : SS3I_binop_rm_int_8 <0x04, "pmaddubsw",
2788 int_x86_ssse3_pmadd_ub_sw,
2789 int_x86_ssse3_pmadd_ub_sw_128>;
2790 defm PMULHRSW : SS3I_binop_rm_int_16<0x0B, "pmulhrsw",
2791 int_x86_ssse3_pmul_hr_sw,
2792 int_x86_ssse3_pmul_hr_sw_128, 1>;
2794 defm PSHUFB : SS3I_binop_rm_int_8 <0x00, "pshufb",
2795 int_x86_ssse3_pshuf_b,
2796 int_x86_ssse3_pshuf_b_128>;
2797 defm PSIGNB : SS3I_binop_rm_int_8 <0x08, "psignb",
2798 int_x86_ssse3_psign_b,
2799 int_x86_ssse3_psign_b_128>;
2800 defm PSIGNW : SS3I_binop_rm_int_16<0x09, "psignw",
2801 int_x86_ssse3_psign_w,
2802 int_x86_ssse3_psign_w_128>;
2803 defm PSIGND : SS3I_binop_rm_int_32<0x0A, "psignd",
2804 int_x86_ssse3_psign_d,
2805 int_x86_ssse3_psign_d_128>;
2808 // palignr patterns.
2809 let Constraints = "$src1 = $dst" in {
2810 def PALIGNR64rr : SS3AI<0x0F, MRMSrcReg, (outs VR64:$dst),
2811 (ins VR64:$src1, VR64:$src2, i8imm:$src3),
2812 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2814 def PALIGNR64rm : SS3AI<0x0F, MRMSrcMem, (outs VR64:$dst),
2815 (ins VR64:$src1, i64mem:$src2, i8imm:$src3),
2816 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2819 def PALIGNR128rr : SS3AI<0x0F, MRMSrcReg, (outs VR128:$dst),
2820 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
2821 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2823 def PALIGNR128rm : SS3AI<0x0F, MRMSrcMem, (outs VR128:$dst),
2824 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
2825 "palignr\t{$src3, $src2, $dst|$dst, $src2, $src3}",
2829 let AddedComplexity = 5 in {
2831 def : Pat<(v1i64 (palign:$src3 VR64:$src1, VR64:$src2)),
2832 (PALIGNR64rr VR64:$src2, VR64:$src1,
2833 (SHUFFLE_get_palign_imm VR64:$src3))>,
2834 Requires<[HasSSSE3]>;
2835 def : Pat<(v2i32 (palign:$src3 VR64:$src1, VR64:$src2)),
2836 (PALIGNR64rr VR64:$src2, VR64:$src1,
2837 (SHUFFLE_get_palign_imm VR64:$src3))>,
2838 Requires<[HasSSSE3]>;
2839 def : Pat<(v2f32 (palign:$src3 VR64:$src1, VR64:$src2)),
2840 (PALIGNR64rr VR64:$src2, VR64:$src1,
2841 (SHUFFLE_get_palign_imm VR64:$src3))>,
2842 Requires<[HasSSSE3]>;
2843 def : Pat<(v4i16 (palign:$src3 VR64:$src1, VR64:$src2)),
2844 (PALIGNR64rr VR64:$src2, VR64:$src1,
2845 (SHUFFLE_get_palign_imm VR64:$src3))>,
2846 Requires<[HasSSSE3]>;
2847 def : Pat<(v8i8 (palign:$src3 VR64:$src1, VR64:$src2)),
2848 (PALIGNR64rr VR64:$src2, VR64:$src1,
2849 (SHUFFLE_get_palign_imm VR64:$src3))>,
2850 Requires<[HasSSSE3]>;
2852 def : Pat<(v4i32 (palign:$src3 VR128:$src1, VR128:$src2)),
2853 (PALIGNR128rr VR128:$src2, VR128:$src1,
2854 (SHUFFLE_get_palign_imm VR128:$src3))>,
2855 Requires<[HasSSSE3]>;
2856 def : Pat<(v4f32 (palign:$src3 VR128:$src1, VR128:$src2)),
2857 (PALIGNR128rr VR128:$src2, VR128:$src1,
2858 (SHUFFLE_get_palign_imm VR128:$src3))>,
2859 Requires<[HasSSSE3]>;
2860 def : Pat<(v8i16 (palign:$src3 VR128:$src1, VR128:$src2)),
2861 (PALIGNR128rr VR128:$src2, VR128:$src1,
2862 (SHUFFLE_get_palign_imm VR128:$src3))>,
2863 Requires<[HasSSSE3]>;
2864 def : Pat<(v16i8 (palign:$src3 VR128:$src1, VR128:$src2)),
2865 (PALIGNR128rr VR128:$src2, VR128:$src1,
2866 (SHUFFLE_get_palign_imm VR128:$src3))>,
2867 Requires<[HasSSSE3]>;
2870 def : Pat<(X86pshufb VR128:$src, VR128:$mask),
2871 (PSHUFBrr128 VR128:$src, VR128:$mask)>, Requires<[HasSSSE3]>;
2872 def : Pat<(X86pshufb VR128:$src, (bc_v16i8 (memopv2i64 addr:$mask))),
2873 (PSHUFBrm128 VR128:$src, addr:$mask)>, Requires<[HasSSSE3]>;
2875 //===---------------------------------------------------------------------===//
2876 // Non-Instruction Patterns
2877 //===---------------------------------------------------------------------===//
2879 // extload f32 -> f64. This matches load+fextend because we have a hack in
2880 // the isel (PreprocessForFPConvert) that can introduce loads after dag
2882 // Since these loads aren't folded into the fextend, we have to match it
2884 let Predicates = [HasSSE2] in
2885 def : Pat<(fextend (loadf32 addr:$src)),
2886 (CVTSS2SDrm addr:$src)>;
2889 let Predicates = [HasSSE2] in {
2890 def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
2891 def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
2892 def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
2893 def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
2894 def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
2895 def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
2896 def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
2897 def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
2898 def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
2899 def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
2900 def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
2901 def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
2902 def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
2903 def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
2904 def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
2905 def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
2906 def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
2907 def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
2908 def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
2909 def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
2910 def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
2911 def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
2912 def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
2913 def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
2914 def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
2915 def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
2916 def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
2917 def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
2918 def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
2919 def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
2922 // Move scalar to XMM zero-extended
2923 // movd to XMM register zero-extends
2924 let AddedComplexity = 15 in {
2925 // Zeroing a VR128 then do a MOVS{S|D} to the lower bits.
2926 def : Pat<(v2f64 (X86vzmovl (v2f64 (scalar_to_vector FR64:$src)))),
2927 (MOVSDrr (v2f64 (V_SET0PS)), FR64:$src)>;
2928 def : Pat<(v4f32 (X86vzmovl (v4f32 (scalar_to_vector FR32:$src)))),
2929 (MOVSSrr (v4f32 (V_SET0PS)), FR32:$src)>;
2930 def : Pat<(v4f32 (X86vzmovl (v4f32 VR128:$src))),
2931 (MOVSSrr (v4f32 (V_SET0PS)),
2932 (f32 (EXTRACT_SUBREG (v4f32 VR128:$src), sub_ss)))>;
2933 def : Pat<(v4i32 (X86vzmovl (v4i32 VR128:$src))),
2934 (MOVSSrr (v4i32 (V_SET0PI)),
2935 (EXTRACT_SUBREG (v4i32 VR128:$src), sub_ss))>;
2938 // Splat v2f64 / v2i64
2939 let AddedComplexity = 10 in {
2940 def : Pat<(splat_lo (v2f64 VR128:$src), (undef)),
2941 (UNPCKLPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2942 def : Pat<(unpckh (v2f64 VR128:$src), (undef)),
2943 (UNPCKHPDrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2944 def : Pat<(splat_lo (v2i64 VR128:$src), (undef)),
2945 (PUNPCKLQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2946 def : Pat<(unpckh (v2i64 VR128:$src), (undef)),
2947 (PUNPCKHQDQrr VR128:$src, VR128:$src)>, Requires<[HasSSE2]>;
2950 // Special unary SHUFPSrri case.
2951 def : Pat<(v4f32 (pshufd:$src3 VR128:$src1, (undef))),
2952 (SHUFPSrri VR128:$src1, VR128:$src1,
2953 (SHUFFLE_get_shuf_imm VR128:$src3))>;
2954 let AddedComplexity = 5 in
2955 def : Pat<(v4f32 (pshufd:$src2 VR128:$src1, (undef))),
2956 (PSHUFDri VR128:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2957 Requires<[HasSSE2]>;
2958 // Special unary SHUFPDrri case.
2959 def : Pat<(v2i64 (pshufd:$src3 VR128:$src1, (undef))),
2960 (SHUFPDrri VR128:$src1, VR128:$src1,
2961 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2962 Requires<[HasSSE2]>;
2963 // Special unary SHUFPDrri case.
2964 def : Pat<(v2f64 (pshufd:$src3 VR128:$src1, (undef))),
2965 (SHUFPDrri VR128:$src1, VR128:$src1,
2966 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2967 Requires<[HasSSE2]>;
2968 // Unary v4f32 shuffle with PSHUF* in order to fold a load.
2969 def : Pat<(pshufd:$src2 (bc_v4i32 (memopv4f32 addr:$src1)), (undef)),
2970 (PSHUFDmi addr:$src1, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2971 Requires<[HasSSE2]>;
2973 // Special binary v4i32 shuffle cases with SHUFPS.
2974 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (v4i32 VR128:$src2))),
2975 (SHUFPSrri VR128:$src1, VR128:$src2,
2976 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2977 Requires<[HasSSE2]>;
2978 def : Pat<(v4i32 (shufp:$src3 VR128:$src1, (bc_v4i32 (memopv2i64 addr:$src2)))),
2979 (SHUFPSrmi VR128:$src1, addr:$src2,
2980 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2981 Requires<[HasSSE2]>;
2982 // Special binary v2i64 shuffle cases using SHUFPDrri.
2983 def : Pat<(v2i64 (shufp:$src3 VR128:$src1, VR128:$src2)),
2984 (SHUFPDrri VR128:$src1, VR128:$src2,
2985 (SHUFFLE_get_shuf_imm VR128:$src3))>,
2986 Requires<[HasSSE2]>;
2988 // vector_shuffle v1, <undef>, <0, 0, 1, 1, ...>
2989 let AddedComplexity = 15 in {
2990 def : Pat<(v4i32 (unpckl_undef:$src2 VR128:$src, (undef))),
2991 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2992 Requires<[OptForSpeed, HasSSE2]>;
2993 def : Pat<(v4f32 (unpckl_undef:$src2 VR128:$src, (undef))),
2994 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
2995 Requires<[OptForSpeed, HasSSE2]>;
2997 let AddedComplexity = 10 in {
2998 def : Pat<(v4f32 (unpckl_undef VR128:$src, (undef))),
2999 (UNPCKLPSrr VR128:$src, VR128:$src)>;
3000 def : Pat<(v16i8 (unpckl_undef VR128:$src, (undef))),
3001 (PUNPCKLBWrr VR128:$src, VR128:$src)>;
3002 def : Pat<(v8i16 (unpckl_undef VR128:$src, (undef))),
3003 (PUNPCKLWDrr VR128:$src, VR128:$src)>;
3004 def : Pat<(v4i32 (unpckl_undef VR128:$src, (undef))),
3005 (PUNPCKLDQrr VR128:$src, VR128:$src)>;
3008 // vector_shuffle v1, <undef>, <2, 2, 3, 3, ...>
3009 let AddedComplexity = 15 in {
3010 def : Pat<(v4i32 (unpckh_undef:$src2 VR128:$src, (undef))),
3011 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3012 Requires<[OptForSpeed, HasSSE2]>;
3013 def : Pat<(v4f32 (unpckh_undef:$src2 VR128:$src, (undef))),
3014 (PSHUFDri VR128:$src, (SHUFFLE_get_shuf_imm VR128:$src2))>,
3015 Requires<[OptForSpeed, HasSSE2]>;
3017 let AddedComplexity = 10 in {
3018 def : Pat<(v4f32 (unpckh_undef VR128:$src, (undef))),
3019 (UNPCKHPSrr VR128:$src, VR128:$src)>;
3020 def : Pat<(v16i8 (unpckh_undef VR128:$src, (undef))),
3021 (PUNPCKHBWrr VR128:$src, VR128:$src)>;
3022 def : Pat<(v8i16 (unpckh_undef VR128:$src, (undef))),
3023 (PUNPCKHWDrr VR128:$src, VR128:$src)>;
3024 def : Pat<(v4i32 (unpckh_undef VR128:$src, (undef))),
3025 (PUNPCKHDQrr VR128:$src, VR128:$src)>;
3028 let AddedComplexity = 20 in {
3029 // vector_shuffle v1, v2 <0, 1, 4, 5> using MOVLHPS
3030 def : Pat<(v4i32 (movlhps VR128:$src1, VR128:$src2)),
3031 (MOVLHPSrr VR128:$src1, VR128:$src2)>;
3033 // vector_shuffle v1, v2 <6, 7, 2, 3> using MOVHLPS
3034 def : Pat<(v4i32 (movhlps VR128:$src1, VR128:$src2)),
3035 (MOVHLPSrr VR128:$src1, VR128:$src2)>;
3037 // vector_shuffle v1, undef <2, ?, ?, ?> using MOVHLPS
3038 def : Pat<(v4f32 (movhlps_undef VR128:$src1, (undef))),
3039 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3040 def : Pat<(v4i32 (movhlps_undef VR128:$src1, (undef))),
3041 (MOVHLPSrr VR128:$src1, VR128:$src1)>;
3044 let AddedComplexity = 20 in {
3045 // vector_shuffle v1, (load v2) <4, 5, 2, 3> using MOVLPS
3046 def : Pat<(v4f32 (movlp VR128:$src1, (load addr:$src2))),
3047 (MOVLPSrm VR128:$src1, addr:$src2)>;
3048 def : Pat<(v2f64 (movlp VR128:$src1, (load addr:$src2))),
3049 (MOVLPDrm VR128:$src1, addr:$src2)>;
3050 def : Pat<(v4i32 (movlp VR128:$src1, (load addr:$src2))),
3051 (MOVLPSrm VR128:$src1, addr:$src2)>;
3052 def : Pat<(v2i64 (movlp VR128:$src1, (load addr:$src2))),
3053 (MOVLPDrm VR128:$src1, addr:$src2)>;
3056 // (store (vector_shuffle (load addr), v2, <4, 5, 2, 3>), addr) using MOVLPS
3057 def : Pat<(store (v4f32 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3058 (MOVLPSmr addr:$src1, VR128:$src2)>;
3059 def : Pat<(store (v2f64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3060 (MOVLPDmr addr:$src1, VR128:$src2)>;
3061 def : Pat<(store (v4i32 (movlp (bc_v4i32 (loadv2i64 addr:$src1)), VR128:$src2)),
3063 (MOVLPSmr addr:$src1, VR128:$src2)>;
3064 def : Pat<(store (v2i64 (movlp (load addr:$src1), VR128:$src2)), addr:$src1),
3065 (MOVLPDmr addr:$src1, VR128:$src2)>;
3067 let AddedComplexity = 15 in {
3068 // Setting the lowest element in the vector.
3069 def : Pat<(v4i32 (movl VR128:$src1, VR128:$src2)),
3070 (MOVSSrr (v4i32 VR128:$src1),
3071 (EXTRACT_SUBREG (v4i32 VR128:$src2), sub_ss))>;
3072 def : Pat<(v2i64 (movl VR128:$src1, VR128:$src2)),
3073 (MOVSDrr (v2i64 VR128:$src1),
3074 (EXTRACT_SUBREG (v2i64 VR128:$src2), sub_sd))>;
3076 // vector_shuffle v1, v2 <4, 5, 2, 3> using movsd
3077 def : Pat<(v4f32 (movlp VR128:$src1, VR128:$src2)),
3078 (MOVSDrr VR128:$src1, (EXTRACT_SUBREG VR128:$src2, sub_sd))>,
3079 Requires<[HasSSE2]>;
3080 def : Pat<(v4i32 (movlp VR128:$src1, VR128:$src2)),
3081 (MOVSDrr VR128:$src1, (EXTRACT_SUBREG VR128:$src2, sub_sd))>,
3082 Requires<[HasSSE2]>;
3085 // vector_shuffle v1, v2 <4, 5, 2, 3> using SHUFPSrri (we prefer movsd, but
3086 // fall back to this for SSE1)
3087 def : Pat<(v4f32 (movlp:$src3 VR128:$src1, (v4f32 VR128:$src2))),
3088 (SHUFPSrri VR128:$src2, VR128:$src1,
3089 (SHUFFLE_get_shuf_imm VR128:$src3))>;
3091 // Set lowest element and zero upper elements.
3092 def : Pat<(v2f64 (X86vzmovl (v2f64 VR128:$src))),
3093 (MOVZPQILo2PQIrr VR128:$src)>, Requires<[HasSSE2]>;
3095 // Some special case pandn patterns.
3096 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3098 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3099 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3101 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3102 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3104 (PANDNrr VR128:$src1, VR128:$src2)>, Requires<[HasSSE2]>;
3106 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v4i32 immAllOnesV))),
3107 (memop addr:$src2))),
3108 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3109 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v8i16 immAllOnesV))),
3110 (memop addr:$src2))),
3111 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3112 def : Pat<(v2i64 (and (xor VR128:$src1, (bc_v2i64 (v16i8 immAllOnesV))),
3113 (memop addr:$src2))),
3114 (PANDNrm VR128:$src1, addr:$src2)>, Requires<[HasSSE2]>;
3116 // vector -> vector casts
3117 def : Pat<(v4f32 (sint_to_fp (v4i32 VR128:$src))),
3118 (Int_CVTDQ2PSrr VR128:$src)>, Requires<[HasSSE2]>;
3119 def : Pat<(v4i32 (fp_to_sint (v4f32 VR128:$src))),
3120 (Int_CVTTPS2DQrr VR128:$src)>, Requires<[HasSSE2]>;
3121 def : Pat<(v2f64 (sint_to_fp (v2i32 VR64:$src))),
3122 (Int_CVTPI2PDrr VR64:$src)>, Requires<[HasSSE2]>;
3123 def : Pat<(v2i32 (fp_to_sint (v2f64 VR128:$src))),
3124 (Int_CVTTPD2PIrr VR128:$src)>, Requires<[HasSSE2]>;
3126 // Use movaps / movups for SSE integer load / store (one byte shorter).
3127 def : Pat<(alignedloadv4i32 addr:$src),
3128 (MOVAPSrm addr:$src)>;
3129 def : Pat<(loadv4i32 addr:$src),
3130 (MOVUPSrm addr:$src)>;
3131 def : Pat<(alignedloadv2i64 addr:$src),
3132 (MOVAPSrm addr:$src)>;
3133 def : Pat<(loadv2i64 addr:$src),
3134 (MOVUPSrm addr:$src)>;
3136 def : Pat<(alignedstore (v2i64 VR128:$src), addr:$dst),
3137 (MOVAPSmr addr:$dst, VR128:$src)>;
3138 def : Pat<(alignedstore (v4i32 VR128:$src), addr:$dst),
3139 (MOVAPSmr addr:$dst, VR128:$src)>;
3140 def : Pat<(alignedstore (v8i16 VR128:$src), addr:$dst),
3141 (MOVAPSmr addr:$dst, VR128:$src)>;
3142 def : Pat<(alignedstore (v16i8 VR128:$src), addr:$dst),
3143 (MOVAPSmr addr:$dst, VR128:$src)>;
3144 def : Pat<(store (v2i64 VR128:$src), addr:$dst),
3145 (MOVUPSmr addr:$dst, VR128:$src)>;
3146 def : Pat<(store (v4i32 VR128:$src), addr:$dst),
3147 (MOVUPSmr addr:$dst, VR128:$src)>;
3148 def : Pat<(store (v8i16 VR128:$src), addr:$dst),
3149 (MOVUPSmr addr:$dst, VR128:$src)>;
3150 def : Pat<(store (v16i8 VR128:$src), addr:$dst),
3151 (MOVUPSmr addr:$dst, VR128:$src)>;
3153 //===----------------------------------------------------------------------===//
3154 // SSE4.1 Instructions
3155 //===----------------------------------------------------------------------===//
3157 multiclass sse41_fp_unop_rm<bits<8> opcps, bits<8> opcpd,
3160 Intrinsic V2F64Int> {
3161 // Intrinsic operation, reg.
3162 // Vector intrinsic operation, reg
3163 def PSr_Int : SS4AIi8<opcps, MRMSrcReg,
3164 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3165 !strconcat(OpcodeStr,
3166 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3167 [(set VR128:$dst, (V4F32Int VR128:$src1, imm:$src2))]>,
3170 // Vector intrinsic operation, mem
3171 def PSm_Int : Ii8<opcps, MRMSrcMem,
3172 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3173 !strconcat(OpcodeStr,
3174 "ps\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3176 (V4F32Int (memopv4f32 addr:$src1),imm:$src2))]>,
3178 Requires<[HasSSE41]>;
3180 // Vector intrinsic operation, reg
3181 def PDr_Int : SS4AIi8<opcpd, MRMSrcReg,
3182 (outs VR128:$dst), (ins VR128:$src1, i32i8imm:$src2),
3183 !strconcat(OpcodeStr,
3184 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3185 [(set VR128:$dst, (V2F64Int VR128:$src1, imm:$src2))]>,
3188 // Vector intrinsic operation, mem
3189 def PDm_Int : SS4AIi8<opcpd, MRMSrcMem,
3190 (outs VR128:$dst), (ins f128mem:$src1, i32i8imm:$src2),
3191 !strconcat(OpcodeStr,
3192 "pd\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3194 (V2F64Int (memopv2f64 addr:$src1),imm:$src2))]>,
3198 let Constraints = "$src1 = $dst" in {
3199 multiclass sse41_fp_binop_rm<bits<8> opcss, bits<8> opcsd,
3203 // Intrinsic operation, reg.
3204 def SSr_Int : SS4AIi8<opcss, MRMSrcReg,
3206 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3207 !strconcat(OpcodeStr,
3208 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3210 (F32Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3213 // Intrinsic operation, mem.
3214 def SSm_Int : SS4AIi8<opcss, MRMSrcMem,
3216 (ins VR128:$src1, ssmem:$src2, i32i8imm:$src3),
3217 !strconcat(OpcodeStr,
3218 "ss\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3220 (F32Int VR128:$src1, sse_load_f32:$src2, imm:$src3))]>,
3223 // Intrinsic operation, reg.
3224 def SDr_Int : SS4AIi8<opcsd, MRMSrcReg,
3226 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3227 !strconcat(OpcodeStr,
3228 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3230 (F64Int VR128:$src1, VR128:$src2, imm:$src3))]>,
3233 // Intrinsic operation, mem.
3234 def SDm_Int : SS4AIi8<opcsd, MRMSrcMem,
3236 (ins VR128:$src1, sdmem:$src2, i32i8imm:$src3),
3237 !strconcat(OpcodeStr,
3238 "sd\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3240 (F64Int VR128:$src1, sse_load_f64:$src2, imm:$src3))]>,
3245 // FP round - roundss, roundps, roundsd, roundpd
3246 defm ROUND : sse41_fp_unop_rm<0x08, 0x09, "round",
3247 int_x86_sse41_round_ps, int_x86_sse41_round_pd>;
3248 defm ROUND : sse41_fp_binop_rm<0x0A, 0x0B, "round",
3249 int_x86_sse41_round_ss, int_x86_sse41_round_sd>;
3251 // SS41I_unop_rm_int_v16 - SSE 4.1 unary operator whose type is v8i16.
3252 multiclass SS41I_unop_rm_int_v16<bits<8> opc, string OpcodeStr,
3253 Intrinsic IntId128> {
3254 def rr128 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3256 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3257 [(set VR128:$dst, (IntId128 VR128:$src))]>, OpSize;
3258 def rm128 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3260 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3263 (bitconvert (memopv8i16 addr:$src))))]>, OpSize;
3266 defm PHMINPOSUW : SS41I_unop_rm_int_v16 <0x41, "phminposuw",
3267 int_x86_sse41_phminposuw>;
3269 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3270 let Constraints = "$src1 = $dst" in {
3271 multiclass SS41I_binop_rm_int<bits<8> opc, string OpcodeStr,
3272 Intrinsic IntId128, bit Commutable = 0> {
3273 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3274 (ins VR128:$src1, VR128:$src2),
3275 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3276 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3278 let isCommutable = Commutable;
3280 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3281 (ins VR128:$src1, i128mem:$src2),
3282 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3284 (IntId128 VR128:$src1,
3285 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3289 defm PCMPEQQ : SS41I_binop_rm_int<0x29, "pcmpeqq",
3290 int_x86_sse41_pcmpeqq, 1>;
3291 defm PACKUSDW : SS41I_binop_rm_int<0x2B, "packusdw",
3292 int_x86_sse41_packusdw, 0>;
3293 defm PMINSB : SS41I_binop_rm_int<0x38, "pminsb",
3294 int_x86_sse41_pminsb, 1>;
3295 defm PMINSD : SS41I_binop_rm_int<0x39, "pminsd",
3296 int_x86_sse41_pminsd, 1>;
3297 defm PMINUD : SS41I_binop_rm_int<0x3B, "pminud",
3298 int_x86_sse41_pminud, 1>;
3299 defm PMINUW : SS41I_binop_rm_int<0x3A, "pminuw",
3300 int_x86_sse41_pminuw, 1>;
3301 defm PMAXSB : SS41I_binop_rm_int<0x3C, "pmaxsb",
3302 int_x86_sse41_pmaxsb, 1>;
3303 defm PMAXSD : SS41I_binop_rm_int<0x3D, "pmaxsd",
3304 int_x86_sse41_pmaxsd, 1>;
3305 defm PMAXUD : SS41I_binop_rm_int<0x3F, "pmaxud",
3306 int_x86_sse41_pmaxud, 1>;
3307 defm PMAXUW : SS41I_binop_rm_int<0x3E, "pmaxuw",
3308 int_x86_sse41_pmaxuw, 1>;
3310 defm PMULDQ : SS41I_binop_rm_int<0x28, "pmuldq", int_x86_sse41_pmuldq, 1>;
3312 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, VR128:$src2)),
3313 (PCMPEQQrr VR128:$src1, VR128:$src2)>;
3314 def : Pat<(v2i64 (X86pcmpeqq VR128:$src1, (memop addr:$src2))),
3315 (PCMPEQQrm VR128:$src1, addr:$src2)>;
3317 /// SS41I_binop_rm_int - Simple SSE 4.1 binary operator
3318 let Constraints = "$src1 = $dst" in {
3319 multiclass SS41I_binop_patint<bits<8> opc, string OpcodeStr, ValueType OpVT,
3320 SDNode OpNode, Intrinsic IntId128,
3321 bit Commutable = 0> {
3322 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3323 (ins VR128:$src1, VR128:$src2),
3324 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3325 [(set VR128:$dst, (OpNode (OpVT VR128:$src1),
3326 VR128:$src2))]>, OpSize {
3327 let isCommutable = Commutable;
3329 def rr_int : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3330 (ins VR128:$src1, VR128:$src2),
3331 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3332 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3334 let isCommutable = Commutable;
3336 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3337 (ins VR128:$src1, i128mem:$src2),
3338 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3340 (OpVT (OpNode VR128:$src1, (memop addr:$src2))))]>, OpSize;
3341 def rm_int : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3342 (ins VR128:$src1, i128mem:$src2),
3343 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3345 (IntId128 VR128:$src1, (memop addr:$src2)))]>,
3350 /// SS48I_binop_rm - Simple SSE41 binary operator.
3351 let Constraints = "$src1 = $dst" in {
3352 multiclass SS48I_binop_rm<bits<8> opc, string OpcodeStr, SDNode OpNode,
3353 ValueType OpVT, bit Commutable = 0> {
3354 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3355 (ins VR128:$src1, VR128:$src2),
3356 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3357 [(set VR128:$dst, (OpVT (OpNode VR128:$src1, VR128:$src2)))]>,
3359 let isCommutable = Commutable;
3361 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3362 (ins VR128:$src1, i128mem:$src2),
3363 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3364 [(set VR128:$dst, (OpNode VR128:$src1,
3365 (bc_v4i32 (memopv2i64 addr:$src2))))]>,
3370 defm PMULLD : SS48I_binop_rm<0x40, "pmulld", mul, v4i32, 1>;
3372 /// SS41I_binop_rmi_int - SSE 4.1 binary operator with 8-bit immediate
3373 let Constraints = "$src1 = $dst" in {
3374 multiclass SS41I_binop_rmi_int<bits<8> opc, string OpcodeStr,
3375 Intrinsic IntId128, bit Commutable = 0> {
3376 def rri : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3377 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3378 !strconcat(OpcodeStr,
3379 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3381 (IntId128 VR128:$src1, VR128:$src2, imm:$src3))]>,
3383 let isCommutable = Commutable;
3385 def rmi : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3386 (ins VR128:$src1, i128mem:$src2, i32i8imm:$src3),
3387 !strconcat(OpcodeStr,
3388 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3390 (IntId128 VR128:$src1,
3391 (bitconvert (memopv16i8 addr:$src2)), imm:$src3))]>,
3396 defm BLENDPS : SS41I_binop_rmi_int<0x0C, "blendps",
3397 int_x86_sse41_blendps, 0>;
3398 defm BLENDPD : SS41I_binop_rmi_int<0x0D, "blendpd",
3399 int_x86_sse41_blendpd, 0>;
3400 defm PBLENDW : SS41I_binop_rmi_int<0x0E, "pblendw",
3401 int_x86_sse41_pblendw, 0>;
3402 defm DPPS : SS41I_binop_rmi_int<0x40, "dpps",
3403 int_x86_sse41_dpps, 1>;
3404 defm DPPD : SS41I_binop_rmi_int<0x41, "dppd",
3405 int_x86_sse41_dppd, 1>;
3406 defm MPSADBW : SS41I_binop_rmi_int<0x42, "mpsadbw",
3407 int_x86_sse41_mpsadbw, 0>;
3410 /// SS41I_ternary_int - SSE 4.1 ternary operator
3411 let Uses = [XMM0], Constraints = "$src1 = $dst" in {
3412 multiclass SS41I_ternary_int<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3413 def rr0 : SS48I<opc, MRMSrcReg, (outs VR128:$dst),
3414 (ins VR128:$src1, VR128:$src2),
3415 !strconcat(OpcodeStr,
3416 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3417 [(set VR128:$dst, (IntId VR128:$src1, VR128:$src2, XMM0))]>,
3420 def rm0 : SS48I<opc, MRMSrcMem, (outs VR128:$dst),
3421 (ins VR128:$src1, i128mem:$src2),
3422 !strconcat(OpcodeStr,
3423 "\t{%xmm0, $src2, $dst|$dst, $src2, %xmm0}"),
3426 (bitconvert (memopv16i8 addr:$src2)), XMM0))]>, OpSize;
3430 defm BLENDVPD : SS41I_ternary_int<0x15, "blendvpd", int_x86_sse41_blendvpd>;
3431 defm BLENDVPS : SS41I_ternary_int<0x14, "blendvps", int_x86_sse41_blendvps>;
3432 defm PBLENDVB : SS41I_ternary_int<0x10, "pblendvb", int_x86_sse41_pblendvb>;
3435 multiclass SS41I_binop_rm_int8<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3436 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3437 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3438 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3440 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i64mem:$src),
3441 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3443 (IntId (bitconvert (v2i64 (scalar_to_vector (loadi64 addr:$src))))))]>,
3447 defm PMOVSXBW : SS41I_binop_rm_int8<0x20, "pmovsxbw", int_x86_sse41_pmovsxbw>;
3448 defm PMOVSXWD : SS41I_binop_rm_int8<0x23, "pmovsxwd", int_x86_sse41_pmovsxwd>;
3449 defm PMOVSXDQ : SS41I_binop_rm_int8<0x25, "pmovsxdq", int_x86_sse41_pmovsxdq>;
3450 defm PMOVZXBW : SS41I_binop_rm_int8<0x30, "pmovzxbw", int_x86_sse41_pmovzxbw>;
3451 defm PMOVZXWD : SS41I_binop_rm_int8<0x33, "pmovzxwd", int_x86_sse41_pmovzxwd>;
3452 defm PMOVZXDQ : SS41I_binop_rm_int8<0x35, "pmovzxdq", int_x86_sse41_pmovzxdq>;
3454 // Common patterns involving scalar load.
3455 def : Pat<(int_x86_sse41_pmovsxbw (vzmovl_v2i64 addr:$src)),
3456 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3457 def : Pat<(int_x86_sse41_pmovsxbw (vzload_v2i64 addr:$src)),
3458 (PMOVSXBWrm addr:$src)>, Requires<[HasSSE41]>;
3460 def : Pat<(int_x86_sse41_pmovsxwd (vzmovl_v2i64 addr:$src)),
3461 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3462 def : Pat<(int_x86_sse41_pmovsxwd (vzload_v2i64 addr:$src)),
3463 (PMOVSXWDrm addr:$src)>, Requires<[HasSSE41]>;
3465 def : Pat<(int_x86_sse41_pmovsxdq (vzmovl_v2i64 addr:$src)),
3466 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3467 def : Pat<(int_x86_sse41_pmovsxdq (vzload_v2i64 addr:$src)),
3468 (PMOVSXDQrm addr:$src)>, Requires<[HasSSE41]>;
3470 def : Pat<(int_x86_sse41_pmovzxbw (vzmovl_v2i64 addr:$src)),
3471 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3472 def : Pat<(int_x86_sse41_pmovzxbw (vzload_v2i64 addr:$src)),
3473 (PMOVZXBWrm addr:$src)>, Requires<[HasSSE41]>;
3475 def : Pat<(int_x86_sse41_pmovzxwd (vzmovl_v2i64 addr:$src)),
3476 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3477 def : Pat<(int_x86_sse41_pmovzxwd (vzload_v2i64 addr:$src)),
3478 (PMOVZXWDrm addr:$src)>, Requires<[HasSSE41]>;
3480 def : Pat<(int_x86_sse41_pmovzxdq (vzmovl_v2i64 addr:$src)),
3481 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3482 def : Pat<(int_x86_sse41_pmovzxdq (vzload_v2i64 addr:$src)),
3483 (PMOVZXDQrm addr:$src)>, Requires<[HasSSE41]>;
3486 multiclass SS41I_binop_rm_int4<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3487 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3488 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3489 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3491 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i32mem:$src),
3492 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3494 (IntId (bitconvert (v4i32 (scalar_to_vector (loadi32 addr:$src))))))]>,
3498 defm PMOVSXBD : SS41I_binop_rm_int4<0x21, "pmovsxbd", int_x86_sse41_pmovsxbd>;
3499 defm PMOVSXWQ : SS41I_binop_rm_int4<0x24, "pmovsxwq", int_x86_sse41_pmovsxwq>;
3500 defm PMOVZXBD : SS41I_binop_rm_int4<0x31, "pmovzxbd", int_x86_sse41_pmovzxbd>;
3501 defm PMOVZXWQ : SS41I_binop_rm_int4<0x34, "pmovzxwq", int_x86_sse41_pmovzxwq>;
3503 // Common patterns involving scalar load
3504 def : Pat<(int_x86_sse41_pmovsxbd (vzmovl_v4i32 addr:$src)),
3505 (PMOVSXBDrm addr:$src)>, Requires<[HasSSE41]>;
3506 def : Pat<(int_x86_sse41_pmovsxwq (vzmovl_v4i32 addr:$src)),
3507 (PMOVSXWQrm addr:$src)>, Requires<[HasSSE41]>;
3509 def : Pat<(int_x86_sse41_pmovzxbd (vzmovl_v4i32 addr:$src)),
3510 (PMOVZXBDrm addr:$src)>, Requires<[HasSSE41]>;
3511 def : Pat<(int_x86_sse41_pmovzxwq (vzmovl_v4i32 addr:$src)),
3512 (PMOVZXWQrm addr:$src)>, Requires<[HasSSE41]>;
3515 multiclass SS41I_binop_rm_int2<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
3516 def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),
3517 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3518 [(set VR128:$dst, (IntId VR128:$src))]>, OpSize;
3520 // Expecting a i16 load any extended to i32 value.
3521 def rm : SS48I<opc, MRMSrcMem, (outs VR128:$dst), (ins i16mem:$src),
3522 !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
3523 [(set VR128:$dst, (IntId (bitconvert
3524 (v4i32 (scalar_to_vector (loadi16_anyext addr:$src))))))]>,
3528 defm PMOVSXBQ : SS41I_binop_rm_int2<0x22, "pmovsxbq", int_x86_sse41_pmovsxbq>;
3529 defm PMOVZXBQ : SS41I_binop_rm_int2<0x32, "pmovzxbq", int_x86_sse41_pmovzxbq>;
3531 // Common patterns involving scalar load
3532 def : Pat<(int_x86_sse41_pmovsxbq
3533 (bitconvert (v4i32 (X86vzmovl
3534 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3535 (PMOVSXBQrm addr:$src)>, Requires<[HasSSE41]>;
3537 def : Pat<(int_x86_sse41_pmovzxbq
3538 (bitconvert (v4i32 (X86vzmovl
3539 (v4i32 (scalar_to_vector (loadi32 addr:$src))))))),
3540 (PMOVZXBQrm addr:$src)>, Requires<[HasSSE41]>;
3543 /// SS41I_binop_ext8 - SSE 4.1 extract 8 bits to 32 bit reg or 8 bit mem
3544 multiclass SS41I_extract8<bits<8> opc, string OpcodeStr> {
3545 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3546 (ins VR128:$src1, i32i8imm:$src2),
3547 !strconcat(OpcodeStr,
3548 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3549 [(set GR32:$dst, (X86pextrb (v16i8 VR128:$src1), imm:$src2))]>,
3551 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3552 (ins i8mem:$dst, VR128:$src1, i32i8imm:$src2),
3553 !strconcat(OpcodeStr,
3554 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3557 // There's an AssertZext in the way of writing the store pattern
3558 // (store (i8 (trunc (X86pextrb (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3561 defm PEXTRB : SS41I_extract8<0x14, "pextrb">;
3564 /// SS41I_extract16 - SSE 4.1 extract 16 bits to memory destination
3565 multiclass SS41I_extract16<bits<8> opc, string OpcodeStr> {
3566 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3567 (ins i16mem:$dst, VR128:$src1, i32i8imm:$src2),
3568 !strconcat(OpcodeStr,
3569 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3572 // There's an AssertZext in the way of writing the store pattern
3573 // (store (i16 (trunc (X86pextrw (v16i8 VR128:$src1), imm:$src2))), addr:$dst)
3576 defm PEXTRW : SS41I_extract16<0x15, "pextrw">;
3579 /// SS41I_extract32 - SSE 4.1 extract 32 bits to int reg or memory destination
3580 multiclass SS41I_extract32<bits<8> opc, string OpcodeStr> {
3581 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3582 (ins VR128:$src1, i32i8imm:$src2),
3583 !strconcat(OpcodeStr,
3584 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3586 (extractelt (v4i32 VR128:$src1), imm:$src2))]>, OpSize;
3587 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3588 (ins i32mem:$dst, VR128:$src1, i32i8imm:$src2),
3589 !strconcat(OpcodeStr,
3590 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3591 [(store (extractelt (v4i32 VR128:$src1), imm:$src2),
3592 addr:$dst)]>, OpSize;
3595 defm PEXTRD : SS41I_extract32<0x16, "pextrd">;
3598 /// SS41I_extractf32 - SSE 4.1 extract 32 bits fp value to int reg or memory
3600 multiclass SS41I_extractf32<bits<8> opc, string OpcodeStr> {
3601 def rr : SS4AIi8<opc, MRMDestReg, (outs GR32:$dst),
3602 (ins VR128:$src1, i32i8imm:$src2),
3603 !strconcat(OpcodeStr,
3604 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3606 (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2))]>,
3608 def mr : SS4AIi8<opc, MRMDestMem, (outs),
3609 (ins f32mem:$dst, VR128:$src1, i32i8imm:$src2),
3610 !strconcat(OpcodeStr,
3611 "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
3612 [(store (extractelt (bc_v4i32 (v4f32 VR128:$src1)), imm:$src2),
3613 addr:$dst)]>, OpSize;
3616 defm EXTRACTPS : SS41I_extractf32<0x17, "extractps">;
3618 // Also match an EXTRACTPS store when the store is done as f32 instead of i32.
3619 def : Pat<(store (f32 (bitconvert (extractelt (bc_v4i32 (v4f32 VR128:$src1)),
3622 (EXTRACTPSmr addr:$dst, VR128:$src1, imm:$src2)>,
3623 Requires<[HasSSE41]>;
3625 let Constraints = "$src1 = $dst" in {
3626 multiclass SS41I_insert8<bits<8> opc, string OpcodeStr> {
3627 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3628 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3629 !strconcat(OpcodeStr,
3630 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3632 (X86pinsrb VR128:$src1, GR32:$src2, imm:$src3))]>, OpSize;
3633 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3634 (ins VR128:$src1, i8mem:$src2, i32i8imm:$src3),
3635 !strconcat(OpcodeStr,
3636 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3638 (X86pinsrb VR128:$src1, (extloadi8 addr:$src2),
3639 imm:$src3))]>, OpSize;
3643 defm PINSRB : SS41I_insert8<0x20, "pinsrb">;
3645 let Constraints = "$src1 = $dst" in {
3646 multiclass SS41I_insert32<bits<8> opc, string OpcodeStr> {
3647 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3648 (ins VR128:$src1, GR32:$src2, i32i8imm:$src3),
3649 !strconcat(OpcodeStr,
3650 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3652 (v4i32 (insertelt VR128:$src1, GR32:$src2, imm:$src3)))]>,
3654 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3655 (ins VR128:$src1, i32mem:$src2, i32i8imm:$src3),
3656 !strconcat(OpcodeStr,
3657 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3659 (v4i32 (insertelt VR128:$src1, (loadi32 addr:$src2),
3660 imm:$src3)))]>, OpSize;
3664 defm PINSRD : SS41I_insert32<0x22, "pinsrd">;
3666 // insertps has a few different modes, there's the first two here below which
3667 // are optimized inserts that won't zero arbitrary elements in the destination
3668 // vector. The next one matches the intrinsic and could zero arbitrary elements
3669 // in the target vector.
3670 let Constraints = "$src1 = $dst" in {
3671 multiclass SS41I_insertf32<bits<8> opc, string OpcodeStr> {
3672 def rr : SS4AIi8<opc, MRMSrcReg, (outs VR128:$dst),
3673 (ins VR128:$src1, VR128:$src2, i32i8imm:$src3),
3674 !strconcat(OpcodeStr,
3675 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3677 (X86insrtps VR128:$src1, VR128:$src2, imm:$src3))]>,
3679 def rm : SS4AIi8<opc, MRMSrcMem, (outs VR128:$dst),
3680 (ins VR128:$src1, f32mem:$src2, i32i8imm:$src3),
3681 !strconcat(OpcodeStr,
3682 "\t{$src3, $src2, $dst|$dst, $src2, $src3}"),
3684 (X86insrtps VR128:$src1,
3685 (v4f32 (scalar_to_vector (loadf32 addr:$src2))),
3686 imm:$src3))]>, OpSize;
3690 defm INSERTPS : SS41I_insertf32<0x21, "insertps">;
3692 def : Pat<(int_x86_sse41_insertps VR128:$src1, VR128:$src2, imm:$src3),
3693 (INSERTPSrr VR128:$src1, VR128:$src2, imm:$src3)>;
3695 // ptest instruction we'll lower to this in X86ISelLowering primarily from
3696 // the intel intrinsic that corresponds to this.
3697 let Defs = [EFLAGS] in {
3698 def PTESTrr : SS48I<0x17, MRMSrcReg, (outs), (ins VR128:$src1, VR128:$src2),
3699 "ptest \t{$src2, $src1|$src1, $src2}",
3700 [(set EFLAGS, (X86ptest VR128:$src1, VR128:$src2))]>,
3702 def PTESTrm : SS48I<0x17, MRMSrcMem, (outs), (ins VR128:$src1, i128mem:$src2),
3703 "ptest \t{$src2, $src1|$src1, $src2}",
3704 [(set EFLAGS, (X86ptest VR128:$src1, (load addr:$src2)))]>,
3708 def MOVNTDQArm : SS48I<0x2A, MRMSrcMem, (outs VR128:$dst), (ins i128mem:$src),
3709 "movntdqa\t{$src, $dst|$dst, $src}",
3710 [(set VR128:$dst, (int_x86_sse41_movntdqa addr:$src))]>,
3714 //===----------------------------------------------------------------------===//
3715 // SSE4.2 Instructions
3716 //===----------------------------------------------------------------------===//
3718 /// SS42I_binop_rm_int - Simple SSE 4.2 binary operator
3719 let Constraints = "$src1 = $dst" in {
3720 multiclass SS42I_binop_rm_int<bits<8> opc, string OpcodeStr,
3721 Intrinsic IntId128, bit Commutable = 0> {
3722 def rr : SS428I<opc, MRMSrcReg, (outs VR128:$dst),
3723 (ins VR128:$src1, VR128:$src2),
3724 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3725 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3727 let isCommutable = Commutable;
3729 def rm : SS428I<opc, MRMSrcMem, (outs VR128:$dst),
3730 (ins VR128:$src1, i128mem:$src2),
3731 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3733 (IntId128 VR128:$src1,
3734 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3738 defm PCMPGTQ : SS42I_binop_rm_int<0x37, "pcmpgtq", int_x86_sse42_pcmpgtq>;
3740 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, VR128:$src2)),
3741 (PCMPGTQrr VR128:$src1, VR128:$src2)>;
3742 def : Pat<(v2i64 (X86pcmpgtq VR128:$src1, (memop addr:$src2))),
3743 (PCMPGTQrm VR128:$src1, addr:$src2)>;
3745 // crc intrinsic instruction
3746 // This set of instructions are only rm, the only difference is the size
3748 let Constraints = "$src1 = $dst" in {
3749 def CRC32m8 : SS42FI<0xF0, MRMSrcMem, (outs GR32:$dst),
3750 (ins GR32:$src1, i8mem:$src2),
3751 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3753 (int_x86_sse42_crc32_8 GR32:$src1,
3754 (load addr:$src2)))]>;
3755 def CRC32r8 : SS42FI<0xF0, MRMSrcReg, (outs GR32:$dst),
3756 (ins GR32:$src1, GR8:$src2),
3757 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3759 (int_x86_sse42_crc32_8 GR32:$src1, GR8:$src2))]>;
3760 def CRC32m16 : SS42FI<0xF1, MRMSrcMem, (outs GR32:$dst),
3761 (ins GR32:$src1, i16mem:$src2),
3762 "crc32{w} \t{$src2, $src1|$src1, $src2}",
3764 (int_x86_sse42_crc32_16 GR32:$src1,
3765 (load addr:$src2)))]>,
3767 def CRC32r16 : SS42FI<0xF1, MRMSrcReg, (outs GR32:$dst),
3768 (ins GR32:$src1, GR16:$src2),
3769 "crc32{w} \t{$src2, $src1|$src1, $src2}",
3771 (int_x86_sse42_crc32_16 GR32:$src1, GR16:$src2))]>,
3773 def CRC32m32 : SS42FI<0xF1, MRMSrcMem, (outs GR32:$dst),
3774 (ins GR32:$src1, i32mem:$src2),
3775 "crc32{l} \t{$src2, $src1|$src1, $src2}",
3777 (int_x86_sse42_crc32_32 GR32:$src1,
3778 (load addr:$src2)))]>;
3779 def CRC32r32 : SS42FI<0xF1, MRMSrcReg, (outs GR32:$dst),
3780 (ins GR32:$src1, GR32:$src2),
3781 "crc32{l} \t{$src2, $src1|$src1, $src2}",
3783 (int_x86_sse42_crc32_32 GR32:$src1, GR32:$src2))]>;
3784 def CRC64m8 : SS42FI<0xF0, MRMSrcMem, (outs GR64:$dst),
3785 (ins GR64:$src1, i8mem:$src2),
3786 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3788 (int_x86_sse42_crc64_8 GR64:$src1,
3789 (load addr:$src2)))]>,
3791 def CRC64r8 : SS42FI<0xF0, MRMSrcReg, (outs GR64:$dst),
3792 (ins GR64:$src1, GR8:$src2),
3793 "crc32{b} \t{$src2, $src1|$src1, $src2}",
3795 (int_x86_sse42_crc64_8 GR64:$src1, GR8:$src2))]>,
3797 def CRC64m64 : SS42FI<0xF1, MRMSrcMem, (outs GR64:$dst),
3798 (ins GR64:$src1, i64mem:$src2),
3799 "crc32{q} \t{$src2, $src1|$src1, $src2}",
3801 (int_x86_sse42_crc64_64 GR64:$src1,
3802 (load addr:$src2)))]>,
3804 def CRC64r64 : SS42FI<0xF1, MRMSrcReg, (outs GR64:$dst),
3805 (ins GR64:$src1, GR64:$src2),
3806 "crc32{q} \t{$src2, $src1|$src1, $src2}",
3808 (int_x86_sse42_crc64_64 GR64:$src1, GR64:$src2))]>,
3812 // String/text processing instructions.
3813 let Defs = [EFLAGS], usesCustomInserter = 1 in {
3814 def PCMPISTRM128REG : SS42AI<0, Pseudo, (outs VR128:$dst),
3815 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3816 "#PCMPISTRM128rr PSEUDO!",
3817 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1, VR128:$src2,
3818 imm:$src3))]>, OpSize;
3819 def PCMPISTRM128MEM : SS42AI<0, Pseudo, (outs VR128:$dst),
3820 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3821 "#PCMPISTRM128rm PSEUDO!",
3822 [(set VR128:$dst, (int_x86_sse42_pcmpistrm128 VR128:$src1, (load addr:$src2),
3823 imm:$src3))]>, OpSize;
3826 let Defs = [XMM0, EFLAGS] in {
3827 def PCMPISTRM128rr : SS42AI<0x62, MRMSrcReg, (outs),
3828 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3829 "pcmpistrm\t{$src3, $src2, $src1|$src1, $src2, $src3}", []>, OpSize;
3830 def PCMPISTRM128rm : SS42AI<0x62, MRMSrcMem, (outs),
3831 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3832 "pcmpistrm\t{$src3, $src2, $src1|$src1, $src2, $src3}", []>, OpSize;
3835 let Defs = [EFLAGS], Uses = [EAX, EDX], usesCustomInserter = 1 in {
3836 def PCMPESTRM128REG : SS42AI<0, Pseudo, (outs VR128:$dst),
3837 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3838 "#PCMPESTRM128rr PSEUDO!",
3840 (int_x86_sse42_pcmpestrm128
3841 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5))]>, OpSize;
3843 def PCMPESTRM128MEM : SS42AI<0, Pseudo, (outs VR128:$dst),
3844 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3845 "#PCMPESTRM128rm PSEUDO!",
3846 [(set VR128:$dst, (int_x86_sse42_pcmpestrm128
3847 VR128:$src1, EAX, (load addr:$src3), EDX, imm:$src5))]>,
3851 let Defs = [XMM0, EFLAGS], Uses = [EAX, EDX] in {
3852 def PCMPESTRM128rr : SS42AI<0x60, MRMSrcReg, (outs),
3853 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3854 "pcmpestrm\t{$src5, $src3, $src1|$src1, $src3, $src5}", []>, OpSize;
3855 def PCMPESTRM128rm : SS42AI<0x60, MRMSrcMem, (outs),
3856 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3857 "pcmpestrm\t{$src5, $src3, $src1|$src1, $src3, $src5}", []>, OpSize;
3860 let Defs = [ECX, EFLAGS] in {
3861 multiclass SS42AI_pcmpistri<Intrinsic IntId128> {
3862 def rr : SS42AI<0x63, MRMSrcReg, (outs),
3863 (ins VR128:$src1, VR128:$src2, i8imm:$src3),
3864 "pcmpistri\t{$src3, $src2, $src1|$src1, $src2, $src3}",
3865 [(set ECX, (IntId128 VR128:$src1, VR128:$src2, imm:$src3)),
3866 (implicit EFLAGS)]>, OpSize;
3867 def rm : SS42AI<0x63, MRMSrcMem, (outs),
3868 (ins VR128:$src1, i128mem:$src2, i8imm:$src3),
3869 "pcmpistri\t{$src3, $src2, $src1|$src1, $src2, $src3}",
3870 [(set ECX, (IntId128 VR128:$src1, (load addr:$src2), imm:$src3)),
3871 (implicit EFLAGS)]>, OpSize;
3875 defm PCMPISTRI : SS42AI_pcmpistri<int_x86_sse42_pcmpistri128>;
3876 defm PCMPISTRIA : SS42AI_pcmpistri<int_x86_sse42_pcmpistria128>;
3877 defm PCMPISTRIC : SS42AI_pcmpistri<int_x86_sse42_pcmpistric128>;
3878 defm PCMPISTRIO : SS42AI_pcmpistri<int_x86_sse42_pcmpistrio128>;
3879 defm PCMPISTRIS : SS42AI_pcmpistri<int_x86_sse42_pcmpistris128>;
3880 defm PCMPISTRIZ : SS42AI_pcmpistri<int_x86_sse42_pcmpistriz128>;
3882 let Defs = [ECX, EFLAGS] in {
3883 let Uses = [EAX, EDX] in {
3884 multiclass SS42AI_pcmpestri<Intrinsic IntId128> {
3885 def rr : SS42AI<0x61, MRMSrcReg, (outs),
3886 (ins VR128:$src1, VR128:$src3, i8imm:$src5),
3887 "pcmpestri\t{$src5, $src3, $src1|$src1, $src3, $src5}",
3888 [(set ECX, (IntId128 VR128:$src1, EAX, VR128:$src3, EDX, imm:$src5)),
3889 (implicit EFLAGS)]>, OpSize;
3890 def rm : SS42AI<0x61, MRMSrcMem, (outs),
3891 (ins VR128:$src1, i128mem:$src3, i8imm:$src5),
3892 "pcmpestri\t{$src5, $src3, $src1|$src1, $src3, $src5}",
3894 (IntId128 VR128:$src1, EAX, (load addr:$src3), EDX, imm:$src5)),
3895 (implicit EFLAGS)]>, OpSize;
3900 defm PCMPESTRI : SS42AI_pcmpestri<int_x86_sse42_pcmpestri128>;
3901 defm PCMPESTRIA : SS42AI_pcmpestri<int_x86_sse42_pcmpestria128>;
3902 defm PCMPESTRIC : SS42AI_pcmpestri<int_x86_sse42_pcmpestric128>;
3903 defm PCMPESTRIO : SS42AI_pcmpestri<int_x86_sse42_pcmpestrio128>;
3904 defm PCMPESTRIS : SS42AI_pcmpestri<int_x86_sse42_pcmpestris128>;
3905 defm PCMPESTRIZ : SS42AI_pcmpestri<int_x86_sse42_pcmpestriz128>;
3907 //===----------------------------------------------------------------------===//
3908 // AES-NI Instructions
3909 //===----------------------------------------------------------------------===//
3911 let Constraints = "$src1 = $dst" in {
3912 multiclass AESI_binop_rm_int<bits<8> opc, string OpcodeStr,
3913 Intrinsic IntId128, bit Commutable = 0> {
3914 def rr : AES8I<opc, MRMSrcReg, (outs VR128:$dst),
3915 (ins VR128:$src1, VR128:$src2),
3916 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3917 [(set VR128:$dst, (IntId128 VR128:$src1, VR128:$src2))]>,
3919 let isCommutable = Commutable;
3921 def rm : AES8I<opc, MRMSrcMem, (outs VR128:$dst),
3922 (ins VR128:$src1, i128mem:$src2),
3923 !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
3925 (IntId128 VR128:$src1,
3926 (bitconvert (memopv16i8 addr:$src2))))]>, OpSize;
3930 defm AESENC : AESI_binop_rm_int<0xDC, "aesenc",
3931 int_x86_aesni_aesenc>;
3932 defm AESENCLAST : AESI_binop_rm_int<0xDD, "aesenclast",
3933 int_x86_aesni_aesenclast>;
3934 defm AESDEC : AESI_binop_rm_int<0xDE, "aesdec",
3935 int_x86_aesni_aesdec>;
3936 defm AESDECLAST : AESI_binop_rm_int<0xDF, "aesdeclast",
3937 int_x86_aesni_aesdeclast>;
3939 def : Pat<(v2i64 (int_x86_aesni_aesenc VR128:$src1, VR128:$src2)),
3940 (AESENCrr VR128:$src1, VR128:$src2)>;
3941 def : Pat<(v2i64 (int_x86_aesni_aesenc VR128:$src1, (memop addr:$src2))),
3942 (AESENCrm VR128:$src1, addr:$src2)>;
3943 def : Pat<(v2i64 (int_x86_aesni_aesenclast VR128:$src1, VR128:$src2)),
3944 (AESENCLASTrr VR128:$src1, VR128:$src2)>;
3945 def : Pat<(v2i64 (int_x86_aesni_aesenclast VR128:$src1, (memop addr:$src2))),
3946 (AESENCLASTrm VR128:$src1, addr:$src2)>;
3947 def : Pat<(v2i64 (int_x86_aesni_aesdec VR128:$src1, VR128:$src2)),
3948 (AESDECrr VR128:$src1, VR128:$src2)>;
3949 def : Pat<(v2i64 (int_x86_aesni_aesdec VR128:$src1, (memop addr:$src2))),
3950 (AESDECrm VR128:$src1, addr:$src2)>;
3951 def : Pat<(v2i64 (int_x86_aesni_aesdeclast VR128:$src1, VR128:$src2)),
3952 (AESDECLASTrr VR128:$src1, VR128:$src2)>;
3953 def : Pat<(v2i64 (int_x86_aesni_aesdeclast VR128:$src1, (memop addr:$src2))),
3954 (AESDECLASTrm VR128:$src1, addr:$src2)>;
3956 def AESIMCrr : AES8I<0xDB, MRMSrcReg, (outs VR128:$dst),
3958 "aesimc\t{$src1, $dst|$dst, $src1}",
3960 (int_x86_aesni_aesimc VR128:$src1))]>,
3963 def AESIMCrm : AES8I<0xDB, MRMSrcMem, (outs VR128:$dst),
3964 (ins i128mem:$src1),
3965 "aesimc\t{$src1, $dst|$dst, $src1}",
3967 (int_x86_aesni_aesimc (bitconvert (memopv2i64 addr:$src1))))]>,
3970 def AESKEYGENASSIST128rr : AESAI<0xDF, MRMSrcReg, (outs VR128:$dst),
3971 (ins VR128:$src1, i8imm:$src2),
3972 "aeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
3974 (int_x86_aesni_aeskeygenassist VR128:$src1, imm:$src2))]>,
3976 def AESKEYGENASSIST128rm : AESAI<0xDF, MRMSrcMem, (outs VR128:$dst),
3977 (ins i128mem:$src1, i8imm:$src2),
3978 "aeskeygenassist\t{$src2, $src1, $dst|$dst, $src1, $src2}",
3980 (int_x86_aesni_aeskeygenassist (bitconvert (memopv2i64 addr:$src1)),